bump mem allocation a bit

12 months ago · 1c4a244291
parent 936dcd2bfc
commit 1c4a244291
1 changed files with 1 additions and 1 deletions
--- a/gpt4all-backend/bert.cpp
+++ b/gpt4all-backend/bert.cpp
@ -871,7 +871,7 @@ struct bert_ctx * bert_load_from_file(const char *fname)

        // TODO: Max tokens should be a param?
        int32_t N = new_bert->model.hparams.n_max_tokens;
-        new_bert->mem_per_input = 1.9 * (new_bert->mem_per_token * N); // add 10% to account for ggml object overhead
+        new_bert->mem_per_input = 2.2 * (new_bert->mem_per_token * N); // add 10% to account for ggml object overhead

    }
 #if defined(DEBUG_BERT)