diff --git a/benchmarks/diffusion/diffusion_benchmark_serving.py b/benchmarks/diffusion/diffusion_benchmark_serving.py
index 878159b81..bf08b8c2b 100644
--- a/benchmarks/diffusion/diffusion_benchmark_serving.py
+++ b/benchmarks/diffusion/diffusion_benchmark_serving.py
@@ -107,6 +107,9 @@ from tqdm.asyncio import tqdm
 
 logger = logging.getLogger(__name__)
 
+_STAGE_METRICS_ENDPOINTS = {"/v1/chat/completions"}
+_RETURN_STAGE_METRICS_FIELD = "return_stage_metrics"
+
 
 class BaseDataset(ABC):
     def __init__(self, args, api_url: str, model: str):
@@ -1163,6 +1166,10 @@ async def benchmark(args):
     requests_list = dataset.get_requests()
     print(f"Prepared {len(requests_list)} requests from {args.dataset} dataset.")
 
+    if args.return_stage_metrics and args.endpoint in _STAGE_METRICS_ENDPOINTS:
+        for req in requests_list:
+            req.extra_body.setdefault(_RETURN_STAGE_METRICS_FIELD, True)
+
     if args.endpoint == "/v1/images/edits":
         for req in requests_list:
             req.default_bot_task = args.bot_task
@@ -1443,6 +1450,11 @@ if __name__ == "__main__":
         default="think",
         help=("bot_task form field for --endpoint /v1/images/edits (think, recaption, think_recaption, vanilla)."),
     )
+    parser.add_argument(
+        "--return-stage-metrics",
+        action="store_true",
+        help="Request stage duration metrics from endpoints that support return_stage_metrics.",
+    )
 
     args = parser.parse_args()
     asyncio.run(benchmark(args))
diff --git a/tests/dfx/perf/scripts/run_diffusion_benchmark.py b/tests/dfx/perf/scripts/run_diffusion_benchmark.py
index 62ce59606..98462743d 100644
--- a/tests/dfx/perf/scripts/run_diffusion_benchmark.py
+++ b/tests/dfx/perf/scripts/run_diffusion_benchmark.py
@@ -556,6 +556,10 @@ def benchmark_params(request, diffusion_server):
 # ---------------------------------------------------------------------------
 
 
+_STAGE_METRICS_ENDPOINTS = {"/v1/chat/completions"}
+_DIFFUSION_PIPELINE_PROFILER_ARG = "enable-diffusion-pipeline-profiler"
+
+
 def run_benchmark(
     host: str,
     port: int,
@@ -608,6 +612,11 @@ def run_benchmark(
         str(tmp_result_file),
     ]
 
+    serve_args_dict = (server_cfg or {}).get("serve_args_dict")
+    profiler_enabled = isinstance(serve_args_dict, dict) and bool(serve_args_dict.get(_DIFFUSION_PIPELINE_PROFILER_ARG))
+    if endpoint in _STAGE_METRICS_ENDPOINTS and profiler_enabled:
+        cmd.append("--return-stage-metrics")
+
     for key, value in params.items():
         if key in exclude_keys or value is None:
             continue