Skip to content

LLM inference in C/C++

C++ ggml

abort if we see a multi buffer (#25276)

Eve3e5036fbfb16
diff --git a/ggml/src/ggml-backend-meta.cpp b/ggml/src/ggml-backend-meta.cpp
index 7bd3291648..1f29ec8671 100644
--- a/ggml/src/ggml-backend-meta.cpp
+++ b/ggml/src/ggml-backend-meta.cpp
@@ -1144,6 +1144,11 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor_impl(ggml_backend_m
ggml_context * simple_ctx = stc.ctxs[j].get();
ggml_backend_buffer_t simple_buf = buf_ctx->bufs[j].get();
+ if ((simple_buf != nullptr) && ggml_backend_buffer_is_multi_buffer(simple_buf)) {
+ // see https://github.com/ggml-org/llama.cpp/issues/22197
+ GGML_ABORT("multi buffers are not supported by the meta backend");
+ }
+
if (split_dim >= 0 && split_dim < GGML_MAX_DIMS) {
// TODO: the following assert fails for llama-parallel even though the results are correct:
// GGML_ASSERT(ggml_is_contiguously_allocated(tensor));