diff --git a/vllm-llama-head_dim.patch b/vllm-llama-head_dim.patch
new file mode 100644
index 0000000..4aed769
--- /dev/null
+++ b/vllm-llama-head_dim.patch
@@ -0,0 +1,13 @@
+--- lib64/python3/dist-packages/vllm/model_executor/models/llama.py
++++ lib64/python3/dist-packages/vllm/model_executor/models/llama.py
+@@ -128,8 +128,8 @@
+             assert tp_size % self.total_num_kv_heads == 0
+         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+         # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+-        self.head_dim = getattr(config, "head_dim",
+-                                self.hidden_size // self.total_num_heads)
++        def_head_dim = self.hidden_size // self.total_num_heads
++        self.head_dim = getattr(config, "head_dim", def_head_dim) or def_head_dim
+         # Phi models introduced a partial_rotary_factor parameter in the config
+         partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
+         self.rotary_dim = int(partial_rotary_factor * self.head_dim)