diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 60199d6d9..4cb04097e 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -1968,12 +1968,20 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) { if (buf == nullptr) { throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); } - // a multi_buffer has no base to lock, its parts are usually pinned memory (host buffer types with a max size) - if (use_mlock && ggml_backend_buffer_is_host(buf) && ggml_backend_buffer_get_base(buf) != nullptr) { - pimpl->mlock_bufs.emplace_back(new llama_mlock); - auto & mlock_buf = pimpl->mlock_bufs.back(); - mlock_buf->init (ggml_backend_buffer_get_base(buf)); - mlock_buf->grow_to(ggml_backend_buffer_get_size(buf)); + // lock the buffers that hold the tensors: a multi_buffer has no base, and its parts are pageable memory when pinning fails + if (use_mlock && !ml.no_alloc && ggml_backend_buffer_is_host(buf)) { + std::vector lock_bufs; + for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) { + if (t->buffer != nullptr && std::find(lock_bufs.begin(), lock_bufs.end(), t->buffer) == lock_bufs.end()) { + lock_bufs.push_back(t->buffer); + } + } + for (ggml_backend_buffer_t lock_buf : lock_bufs) { + pimpl->mlock_bufs.emplace_back(new llama_mlock); + auto & mlock_buf = pimpl->mlock_bufs.back(); + mlock_buf->init (ggml_backend_buffer_get_base(lock_buf)); + mlock_buf->grow_to(ggml_backend_buffer_get_size(lock_buf)); + } } bufs.emplace_back(buf); }