ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-28 02:11:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reenable OpenMP in scheduler async

Browse Source 
Seems solid up to 4 GPUs. It did hang with --max-gpu 6.
This commit is contained in:
Kawrakow
2026-01-15 14:20:24 +00:00
parent d6e5fb00d6
commit 7878553ad6
1 changed files with 13 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
ggml/src/ggml-backend.cpp
View File

@@ -2180,11 +2180,20 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
}
}
if (false && !has_cpu_work) {
int first_reduce = -1;
for (int i = 0; i < sched->n_splits; i++) {
auto split = &sched->splits[i];
if (split->graph.n_nodes == 1 && split->graph.nodes[0]->op == GGML_OP_REDUCE) {
first_reduce = split->backend_id;
break;
}
}
if (!has_cpu_work) {
#pragma omp parallel num_threads(sched->n_backends)
{
int last_reduce = -1;
int last_reduce = first_reduce;
int ith = omp_get_thread_num();
struct ggml_backend_sched_split * splits = sched->splits;
@@ -2217,6 +2226,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
if (ith == split_backend_id) {
sched->statuses[ith] = ggml_backend_sched_eval(sched, split_backend, split);
if (split->n_inputs > 0 && !sched->own_cpy[split_backend_id]) {
sched->needs_sync[split_backend_id] = true;
} else {
@@ -2226,7 +2237,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
}
}
sched->statuses[ith] = ggml_backend_sched_eval(sched, split_backend, split);
}
if (split->graph.nodes[0]->op == GGML_OP_REDUCE) {
@@ -2246,14 +2256,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
#endif
if (!work_done) {
int first_reduce = -1;
for (int i = 0; i < sched->n_splits; i++) {
auto split = &sched->splits[i];
if (split->graph.n_nodes == 1 && split->graph.nodes[0]->op == GGML_OP_REDUCE) {
first_reduce = split->backend_id;
break;
}
}
std::barrier barrier(sched->n_backends);
auto compute = [sched, &barrier, first_reduce] (int ith) {