Support pipeline parallel in V1 Engine (#1700)

### What this PR does / why we need it?
This patch supports pipeline parallel in V1 Engine

### Does this PR introduce _any_ user-facing change?
Yes, users can run PP in V1

### How was this patch tested?
Manully test














- vLLM version: v0.9.2
- vLLM main:
31d5c1797f

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

vllm_ascend/worker/worker_v1.py

@@ -28,8 +28,10 @@ from vllm.config import VllmConfig
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment)
 from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
+from vllm.distributed.parallel_state import get_pp_group, get_tp_group
 from vllm.logger import logger
 from vllm.lora.request import LoRARequest
+from vllm.sequence import IntermediateTensors
 from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, GiB_bytes
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
@@ -206,7 +208,22 @@ class NPUWorker(WorkerBase):
         self,
         scheduler_output: "SchedulerOutput",
     ) -> Optional[ModelRunnerOutput]:
-        output = self.model_runner.execute_model(scheduler_output)
+        intermediate_tensors = None
+        if not get_pp_group().is_first_rank:
+            intermediate_tensors = IntermediateTensors(
+                get_pp_group().recv_tensor_dict(
+                    all_gather_group=get_tp_group()))
+
+        output = self.model_runner.execute_model(scheduler_output,
+                                                 intermediate_tensors)
+        parallel_config = self.vllm_config.parallel_config
+        if parallel_config.distributed_executor_backend != "external_launcher" \
+            and not get_pp_group().is_last_rank:
+            assert isinstance(output, IntermediateTensors)
+            get_pp_group().send_tensor_dict(output.tensors,
+                                            all_gather_group=get_tp_group())
+            return None
+        assert isinstance(output, ModelRunnerOutput)
         return output if self.is_driver_worker else None
 
     def load_model(self) -> None: