diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 16fc9fd7cb4d8..9e4d3b476b675 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -1710,7 +1710,11 @@ class GPUModelRunner(
                     _get_block_table_and_slot_mapping(kv_cache_gid)
                 )
 
-            if self.speculative_config and spec_decode_common_attn_metadata is None:
+            if (
+                self.speculative_config
+                and spec_decode_common_attn_metadata is None
+                and hasattr(self, "drafter")
+            ):
                 if isinstance(self.drafter, EagleProposer):
                     if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
                         spec_decode_common_attn_metadata = cm