diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index 0c600477..bac7345a 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -838,6 +838,14 @@ class NPUModelRunner(GPUModelRunner):
     def _build_attn_state(self, num_reqs, num_scheduled_tokens, num_valid_tokens):
         if np.all(self.input_batch.num_computed_tokens_cpu[:num_reqs] == 0):
             attn_state = AscendAttentionState.PrefillNoCache
+            # If all prompts are shorter than or equal to decode threshold, they should
+            # be treated as SpecDecoding for correct forward path in mla attention backend
+            if (
+                self.speculative_config
+                and self.speculative_config.method == "mtp"
+                and np.all(num_scheduled_tokens <= self.decode_threshold)
+            ):
+                attn_state = AscendAttentionState.SpecDecoding
         # We assume it is the decode stage, where prefill occurs but only one token is not hit in cache.
         elif np.all(num_scheduled_tokens == 1):
             attn_state = AscendAttentionState.DecodeOnly