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

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

vllm/config.py

@@ -285,10 +285,12 @@ class ParallelConfig:
         pipeline_parallel_size: int,
         tensor_parallel_size: int,
         worker_use_ray: bool,
+        max_parallel_loading_workers: Optional[int] = None,
     ) -> None:
         self.pipeline_parallel_size = pipeline_parallel_size
         self.tensor_parallel_size = tensor_parallel_size
         self.worker_use_ray = worker_use_ray
+        self.max_parallel_loading_workers = max_parallel_loading_workers
 
         self.world_size = pipeline_parallel_size * tensor_parallel_size
         if self.world_size > 1:

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,

vllm/engine/llm_engine.py

@@ -143,6 +143,12 @@ class LLMEngine:
             "init_model",
             get_all_outputs=True,
         )
+        self._run_workers(
+            "load_model",
+            get_all_outputs=True,
+            max_concurrent_workers=self.parallel_config.
+            max_parallel_loading_workers,
+        )
 
     def _init_workers_ray(self, placement_group: "PlacementGroup",
                           **ray_remote_kwargs):
@@ -182,6 +188,12 @@ class LLMEngine:
             "init_model",
             get_all_outputs=True,
         )
+        self._run_workers(
+            "load_model",
+            get_all_outputs=True,
+            max_concurrent_workers=self.parallel_config.
+            max_parallel_loading_workers,
+        )
 
     def _verify_args(self) -> None:
         self.model_config.verify_with_parallel_config(self.parallel_config)
@@ -682,16 +694,15 @@ class LLMEngine:
             seq.status = SequenceStatus.FINISHED_STOPPED
             return
 
-    def _run_workers(
+    def _run_workers_in_batch(
         self,
+        workers,
         method: str,
         *args,
-        get_all_outputs: bool = False,
         **kwargs,
-    ) -> Any:
-        """Runs the given method on all workers."""
+    ):
         all_outputs = []
-        for worker in self.workers:
+        for worker in workers:
             if self.parallel_config.worker_use_ray:
                 executor = partial(worker.execute_method.remote, method)
             else:
@@ -699,9 +710,31 @@ class LLMEngine:
 
             output = executor(*args, **kwargs)
             all_outputs.append(output)
-
         if self.parallel_config.worker_use_ray:
             all_outputs = ray.get(all_outputs)
+        return all_outputs
+
+    def _run_workers(
+        self,
+        method: str,
+        *args,
+        get_all_outputs: bool = False,
+        max_concurrent_workers: Optional[int] = None,
+        **kwargs,
+    ) -> Any:
+        """Runs the given method on all workers."""
+        all_outputs = []
+        if max_concurrent_workers:
+            work_groups = [
+                self.workers[i:i + max_concurrent_workers]
+                for i in range(0, len(self.workers), max_concurrent_workers)
+            ]
+        else:
+            work_groups = [self.workers]
+
+        for workers in work_groups:
+            all_outputs.extend(
+                self._run_workers_in_batch(workers, method, *args, **kwargs))
 
         if get_all_outputs:
             return all_outputs

vllm/worker/worker.py

@@ -67,6 +67,8 @@ class Worker:
 
         # Initialize the model.
         set_random_seed(self.model_config.seed)
+
+    def load_model(self):
         self.model = get_model(self.model_config)
 
     @torch.inference_mode()