[Bugfix] Fix distributed bug again in Qwen2.5-VL & Qwen2.5-Omni (vllm-project#16974)

Signed-off-by: fyabc <suyang.fy@alibaba-inc.com>
Signed-off-by: minpeter <kali2005611@gmail.com>

Author: fyabc

vllm/model_executor/models/qwen2_5_vl.py

@@ -198,8 +198,11 @@ def forward(self, x: torch.Tensor):
 
 def all_gather_interleave(local_tensor, hidden_size: int, tp_size: int):
     """All-gather the input tensor interleavely across model parallel group."""
+    import torch.distributed as dist
     gathered_tensors = [torch.zeros_like(local_tensor) for _ in range(tp_size)]
-    parallel_state.get_tp_group().all_gather(gathered_tensors, local_tensor)
+    dist.all_gather(gathered_tensors,
+                    local_tensor,
+                    group=parallel_state.get_tp_group().device_group)
 
     gathered_tensors_split = [
         torch.split(tensor, hidden_size // tp_size, -1)