Skip to content

ggml-cpu: randomly hang forever in ggml_barrier on weak memory model systems #17515

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

ggml-cpu: randomly hang forever in ggml_barrier on weak memory model systems #17515

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
e5aedf2
ggml-cpu: randomly hang forever in ggml_barrier on weak memory model …
DamonFool Nov 26, 2025
2ab38bf
Merge branch 'master' into cpu-hang
DamonFool Nov 28, 2025
e569028
Merge branch 'master' into cpu-hang
DamonFool Dec 1, 2025
c00171d
Merge branch 'master' into cpu-hang
DamonFool Dec 2, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions ggml/src/ggml-cpu/ggml-cpu.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2943,7 +2943,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
// check if thread is active
static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
struct ggml_threadpool * threadpool = state->threadpool;
int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_relaxed);
int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_acquire);
return (state->ith < n_threads);
}

Expand All @@ -2954,7 +2954,7 @@ static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * s
if (state->pending || threadpool->stop || threadpool->pause) { return true; }

// check for new graph/work
int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_acquire);
if (new_graph != state->last_graph) {
state->pending = ggml_graph_compute_thread_active(state);
state->last_graph = new_graph;
Expand Down Expand Up @@ -3061,7 +3061,7 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int
GGML_PRINT_DEBUG("threadpool: n_threads_cur %d n_threads %d\n", threadpool->n_threads_cur, n_threads);

// Update the number of active threads
atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_release);

// Indicate the graph is ready to be processed
// We need the full seq-cst fence here because of the polling threads (used in thread_sync)
Expand Down
Loading