cont : restore padding for n_kv tensor shape

Author: ggerganov

src/llama-kv-cache.cpp

@@ -957,10 +957,14 @@ bool llama_kv_cache::get_has_shift() const {
 uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
     uint32_t result = 0;
 
+    // pad the n_kv value so that the graph remains constant across batches and can be reused
+    // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220)
+    const uint32_t n_pad_cur = std::max(n_pad, 256u);
+
     for (uint32_t s = 0; s < sinfo.n_stream(); ++s) {
         const auto & cells = v_cells[sinfo.strm[s]];
 
-        result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result);
+        result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result);
     }
 
     return result;