diff --git a/vllm_ascend/attention/dsa_v1.py b/vllm_ascend/attention/dsa_v1.py
index 1109d61711d..5b119b0b3bd 100644
--- a/vllm_ascend/attention/dsa_v1.py
+++ b/vllm_ascend/attention/dsa_v1.py
@@ -704,7 +704,7 @@ def build_req_metadata(
                 self.block_table[: self.num_decodes],
                 self.speculative_config.num_speculative_tokens,
                 self.model_config.hf_config.sliding_window,
-                self.block_size,
+                self.storage_block_size,
                 query_start_loc[: self.num_decodes + 1],
                 self.seq_lens[: self.num_decodes],
                 self.num_decode_tokens,
diff --git a/vllm_ascend/models/deepseek_v4/model.py b/vllm_ascend/models/deepseek_v4/model.py
index a5e45171f9e..1556b74abb3 100644
--- a/vllm_ascend/models/deepseek_v4/model.py
+++ b/vllm_ascend/models/deepseek_v4/model.py
@@ -814,6 +814,9 @@ def make_empty_intermediate_tensors(
         # when speculative decoding is enabled; allocating it unconditionally
         # would permanently cost max_num_batched_tokens * hc_dim per rank.
         spec_config = vllm_config.speculative_config
+        needs_mtp_hidden_states = spec_config is not None and (
+            spec_config.use_eagle() or spec_config.uses_draft_model()
+        )
         self._mtp_hidden_buffer = (
             torch.empty(
                 vllm_config.scheduler_config.max_num_batched_tokens,
@@ -821,7 +824,7 @@ def make_empty_intermediate_tensors(
                 dtype=vllm_config.model_config.dtype,
                 device=self.device,
             )
-            if spec_config is not None and spec_config.method == "mtp"
+            if get_pp_group().is_last_rank and needs_mtp_hidden_states
             else None
         )