fix dp issue temperarily

Signed-off-by: ganyi <pleaplusone.gy@gmail.com>

Author: ganyi1996ppo

vllm_ascend/platform.py

@@ -58,7 +58,8 @@ class NPUPlatform(Platform):
     device_type: str = "npu"
     simple_compile_backend: str = "eager"  # Disable torch.compile()
     ray_device_key: str = "NPU"
-    device_control_env_var: str = "ASCEND_RT_VISIBLE_DEVICES"
+    # device_control_env_var: str = "ASCEND_RT_VISIBLE_DEVICES"
+    device_control_env_var: str = ""
     dispatch_key: str = "PrivateUse1"
 
     supported_quantization: list[str] = [ASCEND_QUATIZATION_METHOD]

vllm_ascend/worker/worker_v1.py

@@ -75,6 +75,9 @@ def __init__(
                          distributed_init_method=distributed_init_method,
                          is_driver_worker=is_driver_worker)
         # Try to import mindie_turbo to accelerate vLLM inference.
+        local_dp_rank = self.vllm_config.parallel_config.data_parallel_rank_local
+        world_size = self.vllm_config.parallel_config.world_size
+        self.local_rank_across_dp = local_dp_rank * world_size + self.local_rank
         try_register_lib(
             "mindie_turbo",
             "MindIE Turbo is installed. vLLM inference will be accelerated with MindIE Turbo."
@@ -112,7 +115,7 @@ def wake_up(self, tags: Optional[list[str]] = None) -> None:
 
     def init_device(self):
         if self.device_config.device.type == "npu":
-            self.device = torch.device(f"npu:{self.local_rank}")
+            self.device = torch.device(f"npu:{self.local_rank_across_dp}")
             NPUPlatform.set_device(self.device)
             NPUPlatform.empty_cache()
             self.init_npu_memory = NPUPlatform.mem_get_info()[0]