[Misc] Use RoPE cache for MRoPE (#8396)

2024-09-11 23:13:14 -07:00
parent 295c4730a8
commit 42ffba11ad
1 changed files with 1 additions and 1 deletions
--- a/vllm/model_executor/layers/rotary_embedding.py
+++ b/vllm/model_executor/layers/rotary_embedding.py
@@ -990,7 +990,7 @@ def get_rope(
                base, is_neox_style, dtype, short_factor, long_factor,
                **extra_kwargs)
        elif scaling_type == "mrope":
-            return MRotaryEmbedding(
+            rotary_emb = MRotaryEmbedding(
                head_size,
                rotary_dim,
                max_position,