diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py
index 539e912327d..d9102cb5c9f 100644
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -1511,7 +1511,10 @@ class DeepseekV2Model(nn.Module):
             ("qkv_proj", "v_proj", "v"),
         ]
         # Fused indexer wk + weights_proj (shard 0 = wk, shard 1 = weights_proj)
-        _pending_wk_fp8: dict = {}  # When WK is in FP8, we dequant to BF16 for fusion
+        _pending_wk_fp8 = getattr(self, "_pending_indexer_wk_fp8", None)
+        if _pending_wk_fp8 is None:
+            self._pending_indexer_wk_fp8 = _pending_wk_fp8 = {}
+
         indexer_fused_mapping = [
             ("wk_weights_proj", "wk", 0),
             ("wk_weights_proj", "weights_proj", 1),