fix RAM OOM when load large models in tensor parallel mode. (#1395)

Co-authored-by: ran_lin <rlin@thoughtworks.com>

vllm/engine/arg_utils.py

@@ -22,6 +22,7 @@ class EngineArgs:
     worker_use_ray: bool = False
     pipeline_parallel_size: int = 1
     tensor_parallel_size: int = 1
+    max_parallel_loading_workers: Optional[int] = None
     block_size: int = 16
     swap_space: int = 4  # GiB
     gpu_memory_utilization: float = 0.90
@@ -128,6 +129,12 @@ class EngineArgs:
                             type=int,
                             default=EngineArgs.tensor_parallel_size,
                             help='number of tensor parallel replicas')
+        parser.add_argument(
+            '--max-parallel-loading-workers',
+            type=int,
+            help='load model sequentially in multiple batches, '
+            'to avoid RAM OOM when using tensor '
+            'parallel and large models')
         # KV cache arguments
         parser.add_argument('--block-size',
                             type=int,
@@ -195,7 +202,8 @@ class EngineArgs:
             getattr(model_config.hf_config, 'sliding_window', None))
         parallel_config = ParallelConfig(self.pipeline_parallel_size,
                                          self.tensor_parallel_size,
-                                         self.worker_use_ray)
+                                         self.worker_use_ray,
+                                         self.max_parallel_loading_workers)
         scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
                                            self.max_num_seqs,
                                            model_config.max_model_len,