diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 33be4dccfc710..92822d829a881 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -1641,7 +1641,10 @@ class GPUModelRunner(
         ) -> None:
             attn_group = self.attn_groups[kv_cache_gid][attn_gid]
             builder = attn_group.get_metadata_builder(ubid or 0)
-            cache_key = (kv_cache_groups[kv_cache_gid].kv_cache_spec, type(builder))
+            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
+            if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
+                kv_cache_spec = kv_cache_spec.kv_cache_specs[attn_group.layer_names[0]]
+            cache_key = (kv_cache_spec, type(builder))
 
             cascade_attn_prefix_len = (
                 cascade_attn_prefix_lens[kv_cache_gid][attn_gid]