Make it easy to profile workers with nsight (vllm-project#3162)

Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>

benchmarks/benchmark_latency.py

@@ -26,6 +26,7 @@ def main(args: argparse.Namespace):
         enforce_eager=args.enforce_eager,
         kv_cache_dtype=args.kv_cache_dtype,
         device=args.device,
+        ray_workers_use_nsight=args.ray_workers_use_nsight,
     )
 
     sampling_params = SamplingParams(
@@ -145,5 +146,10 @@ def run_to_completion(profile_dir: Optional[str] = None):
         default="cuda",
         choices=["cuda"],
         help='device type for vLLM execution, supporting CUDA only currently.')
+    parser.add_argument(
+        "--ray-workers-use-nsight",
+        action='store_true',
+        help="If specified, use nsight to profile ray workers",
+    )
     args = parser.parse_args()
     main(args)

vllm/config.py

@@ -382,6 +382,8 @@ class ParallelConfig:
             parallel and large models.
         disable_custom_all_reduce: Disable the custom all-reduce kernel and
             fall back to NCCL.
+        ray_workers_use_nsight: Whether to profile Ray workers with nsight, see
+            https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler.
     """
 
     def __init__(
@@ -391,6 +393,7 @@ def __init__(
         worker_use_ray: bool,
         max_parallel_loading_workers: Optional[int] = None,
         disable_custom_all_reduce: bool = False,
+        ray_workers_use_nsight: bool = False,
     ) -> None:
         self.pipeline_parallel_size = pipeline_parallel_size
         if is_neuron():
@@ -404,6 +407,7 @@ def __init__(
         self.worker_use_ray = worker_use_ray
         self.max_parallel_loading_workers = max_parallel_loading_workers
         self.disable_custom_all_reduce = disable_custom_all_reduce
+        self.ray_workers_use_nsight = ray_workers_use_nsight
 
         self.world_size = pipeline_parallel_size * self.tensor_parallel_size
         # Ray worker is not supported for Neuron backend.
@@ -426,6 +430,9 @@ def _verify_args(self) -> None:
                 logger.info(
                     "Disabled the custom all-reduce kernel because it is not "
                     "supported with pipeline parallelism.")
+        if self.ray_workers_use_nsight and not self.worker_use_ray:
+            raise ValueError("Unable to use nsight profiling unless workers "
+                             "run with Ray.")
 
         # FIXME(woosuk): Fix the stability issues and re-enable the custom
         # all-reduce kernel.

vllm/engine/arg_utils.py

@@ -46,6 +46,7 @@ class EngineArgs:
     lora_dtype = 'auto'
     max_cpu_loras: Optional[int] = None
     device: str = 'auto'
+    ray_workers_use_nsight: bool = False
 
     def __post_init__(self):
         if self.tokenizer is None:
@@ -168,6 +169,10 @@ def add_cli_args(
             help='load model sequentially in multiple batches, '
             'to avoid RAM OOM when using tensor '
             'parallel and large models')
+        parser.add_argument(
+            '--ray-workers-use-nsight',
+            action='store_true',
+            help='If specified, use nsight to profile ray workers')
         # KV cache arguments
         parser.add_argument('--block-size',
                             type=int,
@@ -305,7 +310,8 @@ def create_engine_configs(
                                          self.tensor_parallel_size,
                                          self.worker_use_ray,
                                          self.max_parallel_loading_workers,
-                                         self.disable_custom_all_reduce)
+                                         self.disable_custom_all_reduce,
+                                         self.ray_workers_use_nsight)
         scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
                                            self.max_num_seqs,
                                            model_config.max_model_len,

vllm/engine/llm_engine.py

@@ -124,7 +124,20 @@ def __init__(
             ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0")
             if ray_usage != "1":
                 os.environ["RAY_USAGE_STATS_ENABLED"] = "0"
-            self._init_workers_ray(placement_group)
+            # Pass additional arguments to initialize the worker
+            additional_ray_args = {}
+            if self.parallel_config.ray_workers_use_nsight:
+                logger.info("Configuring Ray workers to use nsight.")
+                additional_ray_args = {
+                    "runtime_env": {
+                        "nsight": {
+                            "t": "cuda,cudnn,cublas",
+                            "o": "'worker_process_%p'",
+                            "cuda-graph-trace": "node",
+                        }
+                    }
+                }
+            self._init_workers_ray(placement_group, **additional_ray_args)
         else:
             self._init_workers()