diff --git a/llama.cpp b/llama.cpp index 365349335..4bd6ffd80 100644 --- a/llama.cpp +++ b/llama.cpp @@ -5815,6 +5815,24 @@ static struct ggml_cgraph * llama_build_graph( GGML_ASSERT(false); } +#if 1 + for (int i = 0; i < result->n_nodes; ++i) { + struct ggml_tensor * node = result->nodes[i]; + if (getenv("SKIP_KQ_ALL")) { + if ( + strcmp(node->name, "KQ") == 0 || + strcmp(node->name, "KQ_scaled") == 0 || + strcmp(node->name, "KQ_masked") == 0 || + strcmp(node->name, "KQ_soft_max") == 0 || + strcmp(node->name, "KQV") == 0 || + false) { + //printf("skipping %s\n", dst->name); + node->op = GGML_OP_NONE; + } + } + } +#endif + return result; }