diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py
index 22d43a4bae18a..4899f5476f955 100644
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -878,11 +878,14 @@ class Indexer(nn.Module):
         )
 
         q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1))
-        # `rotary_emb` is shape-preserving; `q_pe` is already
-        # [num_tokens, n_head, rope_dim].
+        # Note: RoPE (NeoX) can introduce extra leading dimensions during compilation
+        # so we need to reshape back to token-flattened shapes
+        q_pe = q_pe.reshape(-1, self.n_head, self.rope_dim)
+        k_pe = k_pe.reshape(-1, 1, self.rope_dim)
+
         q = torch.cat([q_pe, q_nope], dim=-1)
         # `k_pe` is [num_tokens, 1, rope_dim] (MQA).
-        k = torch.cat([k_pe.squeeze(1), k_nope], dim=-1)
+        k = torch.cat([k_pe.squeeze(-2), k_nope], dim=-1)
 
         # we only quant q here since k quant is fused with cache insertion
         q = q.view(-1, self.head_dim)