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

CUDA: fix padding of GQA to power of 2 in FA (#19115)

Browse Source
This commit is contained in:
Johannes Gäßler
2026-01-26 23:24:58 +01:00
committed by GitHub
parent 8f80d1b254
commit b0311c16d2
3 changed files with 64 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-common.cuh
+25 -18
View File
@@ -629,8 +629,8 @@ static __global__ void flash_attn_mask_to_KV_max(
template<int D, int ncols1, int ncols2> // D == head size
__launch_bounds__(D, 1)
static __global__ void flash_attn_stream_k_fixup(
float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11,
const int nbatch_fa) {
float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03,
const int ne11, const int ne12, const int nbatch_fa) {
constexpr int ncols = ncols1*ncols2;
const int bidx0 = blockIdx.x;
@@ -641,12 +641,14 @@ static __global__ void flash_attn_stream_k_fixup(
const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols);
const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa;
const int iter_j = (ne01 + (ncols1 - 1)) / ncols1;
const int iter_z = (ne02 + (ncols2 - 1)) / ncols2;
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
const int kbc0 = int64_t(bidx0 + 0)*(iter_k*iter_j*iter_z*ne03) / gridDim.x;
const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*iter_z*ne03) / gridDim.x;
const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa;
const int iter_j = (ne01 + (ncols1 - 1)) / ncols1;
const int iter_z_gqa = (gqa_ratio + (ncols2 - 1)) / ncols2;
const int kbc0 = int64_t(bidx0 + 0)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
const bool did_not_have_any_data = kbc0 == kbc0_stop;
const bool wrote_beginning_of_tile = kbc0 % iter_k == 0;
@@ -655,15 +657,19 @@ static __global__ void flash_attn_stream_k_fixup(
return;
}
const int sequence = kbc0 / (iter_k*iter_j*iter_z);
const int zt = (kbc0 - iter_k*iter_j*iter_z*sequence) / (iter_k*iter_j);
const int jt = (kbc0 - iter_k*iter_j*iter_z*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
// z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index
const int sequence = kbc0 /(iter_k*iter_j*iter_z_gqa*ne12);
const int z_KV = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa);
const int zt_gqa = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j);
const int jt = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k;
if (jt*ncols1 + j >= ne01 || zt*ncols2 + c >= ne02) {
const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index.
if (jt*ncols1 + j >= ne01 || zt_gqa*ncols2 + c >= gqa_ratio) {
return;
}
dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + zt*(ncols2*D) + (j*ne02 + c)*D + tid;
dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + zt_Q*D + (j*ne02 + c)*D + tid;
// Load the partial result that needs a fixup:
float dst_val = 0.0f;
@@ -682,7 +688,7 @@ static __global__ void flash_attn_stream_k_fixup(
int bidx = bidx0 - 1;
int kbc_stop = kbc0;
while(true) {
const int kbc = int64_t(bidx)*(iter_k*iter_j*iter_z*ne03) / gridDim.x;
const int kbc = int64_t(bidx)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
if (kbc == kbc_stop) { // Did not have any data.
bidx--;
kbc_stop = kbc;
@@ -883,9 +889,10 @@ void launch_fattn(
}
}
const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
const int ntiles_z = ((Q->ne[2] + ncols2 - 1) / ncols2);
const int ntiles_total = ntiles_x * ntiles_z * Q->ne[3];
const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
const int gqa_ratio = Q->ne[2] / K->ne[2];
const int ntiles_z_gqa = ((gqa_ratio + ncols2 - 1) / ncols2);
const int ntiles_total = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];
// Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
// Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
@@ -960,7 +967,7 @@ void launch_fattn(
blocks_num.x = ntiles_x;
blocks_num.y = parallel_blocks;
blocks_num.z = ntiles_z*Q->ne[3];
blocks_num.z = ntiles_z_gqa*K->ne[2]*Q->ne[3];
if (parallel_blocks > 1) {
dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
@@ -1014,7 +1021,7 @@ void launch_fattn(
flash_attn_stream_k_fixup<DV, ncols1, ncols2>
<<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], nbatch_fa);
((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], K->ne[2], nbatch_fa);
}
} else if (parallel_blocks > 1) {
const dim3 block_dim_combine(DV, 1, 1);