tests : add -INF blocks to the KQ mask in the FA tests (llama/16380)

* tests : add -INF blocks to the KQ mask in the FA tests

* cont : bump -INF block size to 64

Co-authored-by: Jeff Bolz <redacted>
* ggml : prevent division by zero in FA CPU op

---------

Co-authored-by: Jeff Bolz <redacted>

diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 6275c8305a971336a0eda39daacac61a75530382..8e1a2de14f9833163489cfab24eb39b8d3195e95 100644 (file)
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -8135,7 +8135,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         }
 
         // V /= S
-        const float S_inv = 1.0f/S;
+        const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
         ggml_vec_scale_f32(DV, VKQ32, S_inv);
 
         // dst indices