Mr Chunky is such fun!

ggml/src/ggml-cpu/ops.cpp

@@ -10538,7 +10538,7 @@ static void delta_apply_triangular_updates_chunk_f32(float *       attn,
     for (int64_t seq = 0; seq < n_seqs; seq++) {
         for (int i = 0; i < num_chunks; i++) {
             for (int64_t head = 0; head < H_v; head++) {
-                float * attn_ptr = attn + seq * (chunk_size * chunk_size * H_v) + (head * num_chunks + i) * (chunk_size * chunk_size);
+                float * attn_ptr = attn + seq * (chunk_size * chunk_size * num_chunks * H_v) + (head * num_chunks + i) * (chunk_size * chunk_size);
 
                 // Apply triangular updates following the Python reference exactly:
                 // for i in range(1, chunk_size):
@@ -10591,7 +10591,7 @@ static void delta_add_identity_matrix_chunk_f32(float *       matrix,
     for (int64_t seq = 0; seq < n_seqs; seq++) {
         for (int i = 0; i < num_chunks; i++) {
             for (int64_t head = 0; head < H_v; head++) {
-                float * matrix_ptr = matrix + seq * (chunk_size * chunk_size * H_v) +
+                float * matrix_ptr = matrix + seq * (chunk_size * chunk_size * num_chunks * H_v) +
                                      (head * num_chunks + i) * (chunk_size * chunk_size);
                 // Add identity matrix directly
                 for (int64_t i = 0; i < chunk_size; i++) {