furyhawk/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-07-21 10:15:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Turn off DSA on CUDA for now

Browse Source
This commit is contained in:
Kawrakow
2026-07-20 16:18:31 +00:00
parent 2c4fef8a2c
commit faa6528faf
2 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/sweep-bench/sweep-bench.cpp
+4 -4
View File
@@ -191,8 +191,8 @@ int main(int argc, char ** argv) {
// first measure token generation performance at this context size
const auto t_tg_start = ggml_time_us();
printf("======================================== tg_start for n_kv = %u\n", n_kv);
fprintf(stderr, "======================================== tg_start for n_kv = %u\n", n_kv);
//printf("======================================== tg_start for n_kv = %u\n", n_kv);
//fprintf(stderr, "======================================== tg_start for n_kv = %u\n", n_kv);
for (int irep = 0; irep < nrep; ++irep) {
@@ -209,8 +209,8 @@ int main(int argc, char ** argv) {
}
}
printf("======================================== tg_end for n_kv = %u\n", n_kv);
fprintf(stderr, "======================================== tg_end for n_kv = %u\n", n_kv);
//printf("======================================== tg_end for n_kv = %u\n", n_kv);
//fprintf(stderr, "======================================== tg_end for n_kv = %u\n", n_kv);
const auto t_tg_end = ggml_time_us();
ggml/src/ggml-cuda/dsa_attn.cu
+1 -1
View File
@@ -220,7 +220,7 @@ bool ggml_cuda_dsa_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
const ggml_tensor * sink = dst->src[4];
const ggml_tensor * indexer = dst->src[5];
//if (sink) return false; // We do not support sinks at this point
if (sink) return false; // Something still goes wrong with sinks
if (!Q || !K || !V || !mask || !indexer) return false;
if (indexer->ne[0] % 256 != 0) return false; // lazyness to add checks and handle tailes in case of not multiple of 256