Fix KQ mask padding for the Vulkan back-end (#2350)

2026-08-24 18:31:17 +02:00 · 2026-08-24 18:31:17 +02:00 · c574620b12
parent 64109a4d60
commit c574620b12
1 changed files with 4 additions and 0 deletions
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@ -2483,7 +2483,11 @@ extern "C" {
            int                   nk,
            int                   topk_experts);

+#if GGML_USE_VULKAN
+#define GGML_KQ_MASK_PAD 64
+#else
 #define GGML_KQ_MASK_PAD 16
+#endif

    // q:    [n_embd, n_batch,     n_head,    1]
    // k:    [n_embd, n_kv,        n_head_kv, 1]