diff --git a/vllm/v1/core/single_type_kv_cache_manager.py b/vllm/v1/core/single_type_kv_cache_manager.py
index fe9e7a9891941..e6c7150500cc3 100644
--- a/vllm/v1/core/single_type_kv_cache_manager.py
+++ b/vllm/v1/core/single_type_kv_cache_manager.py
@@ -855,7 +855,8 @@ class SinkFullAttentionManager(FullAttentionManager):
             req_blocks = self.req_to_blocks[request_id]
             assert len(req_blocks) == 0
             # Append both sink blocks and hitted prefix cache blocks
-            req_blocks.extend(self.sink_blocks + new_computed_blocks)
+            req_blocks.extend(self.sink_blocks)
+            req_blocks.extend(new_computed_blocks)
             self.num_cached_block[request_id] = len(new_computed_blocks)
         else:
             # A running request. Should not have new computed blocks.
diff --git a/vllm/v1/worker/gpu/attn_utils.py b/vllm/v1/worker/gpu/attn_utils.py
index 6845652688430..604edf3d7ae29 100644
--- a/vllm/v1/worker/gpu/attn_utils.py
+++ b/vllm/v1/worker/gpu/attn_utils.py
@@ -102,8 +102,8 @@ def _reshape_kv_cache(
 
             attn_backend = attn_backends[layer_name]
             if (
-                getattr(kv_cache_spec, "head_size_v", kv_cache_spec.head_size)
-                != kv_cache_spec.head_size
+                hasattr(kv_cache_spec, "head_size_v")
+                and kv_cache_spec.head_size_v != kv_cache_spec.head_size
             ):
                 kwargs = {"head_size_v": kv_cache_spec.head_size_v}
                 stride_kwargs = {"diff_kv": True}
diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 0bf773c84596d..e02e7ee2d4c26 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -5207,8 +5207,8 @@ class GPUModelRunner(
                     kernel_num_blocks = num_blocks * num_blocks_per_kv_block
 
                     if (
-                        getattr(kv_cache_spec, "head_size_v", kv_cache_spec.head_size)
-                        != kv_cache_spec.head_size
+                        hasattr(kv_cache_spec, "head_size_v")
+                        and kv_cache_spec.head_size_v != kv_cache_spec.head_size
                     ):
                         kwargs = {"head_size_v": kv_cache_spec.head_size_v}
                         stride_kwargs = {"diff_kv": True}
diff --git a/vllm/v1/worker/kv_connector_model_runner_mixin.py b/vllm/v1/worker/kv_connector_model_runner_mixin.py
index a31013502c377..a45b83dc788f1 100644
--- a/vllm/v1/worker/kv_connector_model_runner_mixin.py
+++ b/vllm/v1/worker/kv_connector_model_runner_mixin.py
@@ -191,8 +191,8 @@ class KVConnectorModelRunnerMixin:
 
         attn_backend = attn_group.backend
         if (
-            getattr(kv_cache_spec, "head_size_v", kv_cache_spec.head_size)
-            != kv_cache_spec.head_size
+            hasattr(kv_cache_spec, "head_size_v")
+            and kv_cache_spec.head_size_v != kv_cache_spec.head_size
         ):
             kwargs = {"head_size_v": kv_cache_spec.head_size_v}
             stride_kwargs = {"diff_kv": True}
@@ -269,8 +269,8 @@ class KVConnectorModelRunnerMixin:
 
         attn_backend = attn_group.backend
         if (
-            getattr(kv_cache_spec, "head_size_v", kv_cache_spec.head_size)
-            != kv_cache_spec.head_size
+            hasattr(kv_cache_spec, "head_size_v")
+            and kv_cache_spec.head_size_v != kv_cache_spec.head_size
         ):
             kwargs = {"head_size_v": kv_cache_spec.head_size_v}
             stride_kwargs = {"diff_kv": True}