CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-15 11:17:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: faster tile FA, add oob checks, more HSs (#16492)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-10-11 20:54:32 +02:00
committed by GitHub
parent a3cb04744f
commit 11f0af5504
18 changed files with 1358 additions and 784 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
ggml/src/ggml-cuda/fattn-common.cuh
View File

@@ -793,8 +793,6 @@ void launch_fattn(
GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
"the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
ggml_cuda_pool & pool = ctx.pool();
cudaStream_t main_stream = ctx.stream();
const int id = ggml_cuda_get_device();
@@ -878,7 +876,7 @@ void launch_fattn(
// 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
// multiple sequences of possibly different lengths.
if (mask && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
if (mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
const int s31 = mask->nb[1] / sizeof(half2);
const int s33 = mask->nb[3] / sizeof(half2);
@@ -916,8 +914,7 @@ void launch_fattn(
dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
} else {
GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
const int ntiles_KQ = (K->ne[1] + KQ_row_granularity - 1) / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
// parallel_blocks must not be larger than what the tensor size allows:
parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
@@ -946,7 +943,7 @@ void launch_fattn(
blocks_num.x = ntiles_x;
blocks_num.y = parallel_blocks;
blocks_num.z = Q->ne[2]*Q->ne[3];
blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3];
if (parallel_blocks > 1) {
dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));