diff --git a/vllm/patches/deepseek_v4_attention.py b/vllm/patches/deepseek_v4_attention.py
index 35446abd..22317957 100644
--- a/vllm/patches/deepseek_v4_attention.py
+++ b/vllm/patches/deepseek_v4_attention.py
@@ -630,8 +630,12 @@ class DeepseekV4MultiHeadLatentAttentionWrapper(PluggableLayer):
             out.zero_()
             return
 
+        # Apply RoPE to KV (required for correct attention scores)
+        # Q already has RoPE from fused_qnorm_rope_kv_insert_py or _apply_rope_q
+        kv_rope = self._apply_rope_kv(kv, positions)
+
         # Attention using PyTorch SDPA (works on Blackwell)
-        o = full_sdpa_attention(q, kv, self.scale)
+        o = full_sdpa_attention(q, kv_rope, self.scale)
 
         # Write into the output buffer (same shape as original path)
         if self.n_local_heads < self.padded_heads:
@@ -649,6 +653,15 @@ class DeepseekV4MultiHeadLatentAttentionWrapper(PluggableLayer):
         q[:, :, self.nope_head_dim:][:, :, 0::2] = q_rope[:, :, 0::2] * cos_q - q_rope[:, :, 1::2] * sin_q
         q[:, :, self.nope_head_dim:][:, :, 1::2] = q_rope[:, :, 0::2] * sin_q + q_rope[:, :, 1::2] * cos_q
 
+    def _apply_rope_kv(self, kv, positions):
+        """Apply GPT-J RoPE to KV latent and return the result."""
+        from vllm.model_executor.layers.csa_attention import apply_gptj_rope
+        half = self.rope_head_dim // 2
+        cos = self.rotary_emb.cos_sin_cache[positions, :half].to(kv.dtype)
+        sin = self.rotary_emb.cos_sin_cache[positions, half:].to(kv.dtype)
+        kv_rope = apply_gptj_rope(kv.unsqueeze(1), cos, sin, self.nope_head_dim).squeeze(1)
+        return kv_rope
+
     def _fused_qnorm_rope_kv_insert(
         self,
         q: torch.Tensor,