diff --git a/vllm_ascend/attention/mla_v1.py b/vllm_ascend/attention/mla_v1.py
index 189aa38..cb031ae 100644
--- a/vllm_ascend/attention/mla_v1.py
+++ b/vllm_ascend/attention/mla_v1.py
@@ -1077,7 +1077,7 @@ class AscendMLAImpl(MLAAttentionImpl):
             decode_k_nope = None
             assert attn_metadata.decode is not None
             if self.running_in_graph:
-                seq_len = self.rotary_emb.max_position_embeddings
+                seq_len = self.rotary_emb.max_position_embeddings * self.rotary_emb.scaling_factor
                 cos = self.rotary_emb.cos_cached[:seq_len].to(
                     dtype=decode_hs_or_q_c.dtype)
                 sin = self.rotary_emb.sin_cached[:seq_len].to(
@@ -1122,7 +1122,7 @@ class AscendMLAImpl(MLAAttentionImpl):
             prefill_q_nope = prefill_q[..., :self.qk_nope_head_dim]
             if self.torchair_graph_enabled:
                 num_tokens = prefill_hs_or_q_c.shape[0]
-                seq_len = self.rotary_emb.max_position_embeddings
+                seq_len = self.rotary_emb.max_position_embeddings * self.rotary_emb.scaling_factor
                 cos = self.rotary_emb.cos_cached[:seq_len].to(
                     dtype=prefill_q_pe.dtype)
                 sin = self.rotary_emb.sin_cached[:seq_len].to(