diff --git a/requirements/runtime.txt b/requirements/runtime.txt
index f531754b3..05b38c1b7 100644
--- a/requirements/runtime.txt
+++ b/requirements/runtime.txt
@@ -19,5 +19,8 @@ tiktoken
 torch<=2.1.2
 torchvision<=0.16.2
 # Minimum 4.36.0 to support `Cache` data structure used by KV Cache
-transformers>=4.36.0
+# Registering a causal mask in `LlamaModel` is not friendly for very large
+# `max_position_embeddings`. Refer to
+# https://github.com/huggingface/transformers/blob/v4.38.0/src/transformers/models/llama/modeling_llama.py#L921-L923
+transformers>=4.36.0,!=4.38.0,!=4.38.1,!=4.38.2
 transformers_stream_generator