[Bugfix]Fix moe error when sp chunked the hidden_states (#4212)

### What this PR does / why we need it?
Fix moe error when sp chunked the hidden_states by disabling sp by a hacky way

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

---------

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

vllm_ascend/platform.py

@@ -310,7 +310,10 @@ class NPUPlatform(Platform):
 
         if parallel_config and parallel_config.worker_cls == "auto":
             # TODO: this is a tricky way to disable `use_sequence_parallel_moe` in vllm.
+            if vllm_version_is("0.11.0"):
                 os.environ["VLLM_ALL2ALL_BACKEND"] = "flashinfer_all2allv"
+            else:
+                parallel_config.all2all_backend = "flashinfer_all2allv"
             if ascend_config.torchair_graph_config.enabled or ascend_config.enable_shared_expert_dp:
                 parallel_config.worker_cls = "vllm_ascend.torchair.torchair_worker.NPUTorchairWorker"
             else: