ggml-cpu : size set_rows conversion buffer for all threads (#30454)

* ggml-cpu : size set_rows conversion buffer for all threads

SET_ROWS is planned with n_tasks = 1, but every thread runs it and uses its own row of the F16 -> F32 conversion buffer, so threads > 0 wrote past the work buffer when the graph had another multi-threaded op.

* move set_rows to n_task = n_threads group
This commit is contained in:
Aman Gupta authored and GitHub committed 2026-10-14 18:31:53 +03:00
1 parent 62df1815bc
commit 5cdadfc326
1 file changed
+1 -1
+1 -1
View File
@@ -2360,11 +2360,11 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
case GGML_OP_OUT_PROD:
case GGML_OP_SET_ROWS:
{
n_tasks = n_threads;
} break;
case GGML_OP_GET_ROWS:
case GGML_OP_SET_ROWS:
{
// FIXME: get_rows can use additional threads, but the cost of launching additional threads
// decreases performance with GPU offloading