diff --git a/vllm/patches/deepseek_v4_attention.py b/vllm/patches/deepseek_v4_attention.py
index 9da5cd04..35446abd 100644
--- a/vllm/patches/deepseek_v4_attention.py
+++ b/vllm/patches/deepseek_v4_attention.py
@@ -631,7 +631,7 @@ class DeepseekV4MultiHeadLatentAttentionWrapper(PluggableLayer):
             return
 
         # Attention using PyTorch SDPA (works on Blackwell)
-        o = full_sdpa_attention(q, kv, self.softmax_scale)
+        o = full_sdpa_attention(q, kv, self.scale)
 
         # Write into the output buffer (same shape as original path)
         if self.n_local_heads < self.padded_heads: