Skip to content

LLM inference in C/C++

C++ ggml

meta: recreate all subgraphs when the arena is reallocated (#30453)

Igor Matuszewskiecdd29e8b207
diff --git a/ggml/src/ggml-backend-meta.cpp b/ggml/src/ggml-backend-meta.cpp
index 86748b689a..679d5216d2 100644
--- a/ggml/src/ggml-backend-meta.cpp
+++ b/ggml/src/ggml-backend-meta.cpp
@@ -2284,8 +2284,8 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
backend_ctx->ctx.reset(ggml_init(params));
for (size_t j = 0; j < n_backends; j++) {
auto & bcj = backend_ctx->backend_configs[j];
- for (size_t i = 0; i < n_subgraphs; i++) {
- bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+ for (size_t i = 0; i < backend_ctx->max_subgraphs; i++) {
+ bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), backend_ctx->max_nnodes, /*grads =*/ false);
}
}
backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);