1

Author: giuseppe

ggml/src/ggml-vulkan/vulkan-shaders/ssm_scan.comp

@@ -68,6 +68,10 @@ void main() {
         state[j] = s0[s0_base_idx + j * D_STATE + tid];
     }
 
+    if (tid >= D_STATE) {
+        return;
+    }
+
     for (int i = 0; i < int(n_tok); i++) {
         float dt_soft_plus = dt[dt_base_idx + i * stride_dt];
         dt_soft_plus = softplus(dt_soft_plus);
@@ -119,9 +123,8 @@ void main() {
                 barrier();
             }
 
-            y = warp_sdata[tid & ~(SUBGROUP_SIZE - 1)];
-
             if (tid % SUBGROUP_SIZE == 0) {
+                y = warp_sdata[tid & ~(SUBGROUP_SIZE - 1)];
                 const int k = tid / SUBGROUP_SIZE + j * (D_STATE / SUBGROUP_SIZE);
                 d[y_base_idx + uint(i) * uint(stride_y) + uint(k)] = y;
             }