diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index 859f21ece31..c9c94dcc5a0 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -30,7 +30,6 @@
 import numpy.typing as npt
 import torch
 import torch._dynamo.cache_size
-import torch.distributed as dist
 import torch.nn as nn
 from tqdm import tqdm  # type: ignore
 from vllm.attention import AttentionType, get_attn_backend
@@ -587,17 +586,21 @@ def _get_forward_metadata_across_dp(
             self, num_tokens: int, with_prefill: bool,
             enable_dbo: bool) -> tuple[torch.Tensor, bool, bool]:
 
-        # Compose: all_reduce metadata (num_tokens of each rank, with_prefill, enable_dbo)
-        num_tokens_across_dp = torch.zeros(self.dp_size + 2,
-                                           dtype=torch.int32,
-                                           device="cpu")
-        num_tokens_across_dp[self.dp_rank] = num_tokens
-        num_tokens_across_dp[-2] = int(with_prefill)
-        num_tokens_across_dp[-1] = int(not enable_dbo)
-        dist.all_reduce(num_tokens_across_dp, group=get_dp_group().cpu_group)
-        with_prefill = bool(num_tokens_across_dp[-2])
-        enable_dbo = not bool(num_tokens_across_dp[-1])
-        num_tokens_across_dp = num_tokens_across_dp[:-2]
+        local_forward_metadata = torch.tensor(
+            [[num_tokens, with_prefill, enable_dbo]],
+            device="npu",
+            dtype=torch.int32)
+        # use device all_gather
+        global_forward_metadata = get_dp_group().all_gather(
+            local_forward_metadata, dim=0)
+        maybe_padded_num_tokens = global_forward_metadata[:, 0].max().item()
+        num_tokens_across_dp = torch.tensor([maybe_padded_num_tokens] *
+                                            self.dp_size,
+                                            device="cpu",
+                                            dtype=torch.int32)
+        with_prefill = bool(global_forward_metadata[:, 1].any())
+        enable_dbo = bool(global_forward_metadata[:, 2].any())
+
         return num_tokens_across_dp, with_prefill, enable_dbo
 
     def _get_forward_metadata_across_dp_and_pad(