diff --git a/vllm/model_executor/layers/quantization/kv_cache.py b/vllm/model_executor/layers/quantization/kv_cache.py
index c8f9e86a9cc7c..388d3a91228c8 100644
--- a/vllm/model_executor/layers/quantization/kv_cache.py
+++ b/vllm/model_executor/layers/quantization/kv_cache.py
@@ -89,6 +89,7 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
             q_scale = layer.q_scale.to("cpu").tolist()
             if current_platform.is_rocm() and not is_navi():
                 q_scale *= 2
+            layer.calculate_kv_scales = False
         else:
             q_scale = 1.0
         if layer.prob_scale > 0.0: