llama : make pos_bias contiguous for CUDA

2024-06-26 17:46:39 +02:00 · 2024-06-26 17:46:39 +02:00 · c4ded1a8fb
parent bad0cafee9
commit c4ded1a8fb
1 changed files with 3 additions and 0 deletions
--- a/llama.cpp
+++ b/llama.cpp
@ -8104,6 +8104,9 @@ struct llm_build_context {
        pos_bias = ggml_permute(ctx0, pos_bias, 2, 0, 1, 3);
        cb(pos_bias, "pos_bias", -1);

+        pos_bias = ggml_cont(ctx0, pos_bias);
+        cb(pos_bias, "pos_bias", -1);
+
        return pos_bias;
    }