[multihost] Integrate expert parallelism to RayExecutor

Signed-off-by: Lihao Ran <imlihao.ran@gmail.com>

Author: Lumosis

tpu_inference/executors/ray_distributed_executor.py

@@ -366,7 +366,7 @@ def sort_by_driver_then_worker_ip(item: RayWorkerMetaData):
             for pp_rank in range(self.parallel_config.pipeline_parallel_size):
                 self.pp_tp_workers.append([])
                 num_tp_workers = int(
-                    self.parallel_config.tensor_parallel_size //
+                    self.vllm_config.sharding_config.total_devices //
                     num_tpu_per_worker)
                 for tp_rank in range(num_tp_workers):
                     # PP=2, TP=4, num_tpu_per_worker=2