From d257b58e67193780ff8a59ab54b48219b9dc28d2 Mon Sep 17 00:00:00 2001
From: Lianmin Zheng <lianminzheng@gmail.com>
Date: Tue, 4 Aug 2026 05:55:35 -0700
Subject: [PATCH] [Router] Report accelerator count in /v1/loads (#33548)

Co-authored-by: Yinghai Lu <yinghai@meta.com>
---
 python/sglang/srt/entrypoints/v1_loads.py     | 22 +++++++++++-
 .../entrypoints/test_v1_loads_aggregate.py    | 34 ++++++++++++++++---
 2 files changed, 51 insertions(+), 5 deletions(-)

diff --git a/python/sglang/srt/entrypoints/v1_loads.py b/python/sglang/srt/entrypoints/v1_loads.py
index d940da1c3..f7eb046fa 100644
--- a/python/sglang/srt/entrypoints/v1_loads.py
+++ b/python/sglang/srt/entrypoints/v1_loads.py
@@ -38,6 +38,19 @@ def _accelerator_name() -> Optional[str]:
     return get_device_name()
 
 
+@lru_cache(maxsize=1)
+def _num_accelerators_per_dp_rank(
+    tp_size: int,
+    pp_size: int,
+    dp_size: int,
+    enable_dp_attention: bool,
+) -> int:
+    num_accelerators = tp_size * pp_size
+    if enable_dp_attention:
+        num_accelerators //= dp_size
+    return num_accelerators
+
+
 def _get_tokenizer_manager():
     """Dependency to get tokenizer_manager from global state."""
     from sglang.srt.entrypoints.http_server import get_global_state
@@ -95,7 +108,8 @@ async def get_loads(
         format: Response format - 'json' (default) or 'prometheus'
 
     Returns:
-        JSON response with timestamp, version, accelerator, and per-DP-rank loads
+        JSON response with timestamp, version, accelerator metadata, and
+        per-DP-rank loads
     """
     include_list = [s.strip() for s in include.split(",")] if include else None
 
@@ -128,5 +142,11 @@ async def get_loads(
         "timestamp": datetime.now(timezone.utc).isoformat(),
         "version": __version__,
         "accelerator": _accelerator_name(),
+        "num_accelerators": _num_accelerators_per_dp_rank(
+            tokenizer_manager.server_args.tp_size,
+            tokenizer_manager.server_args.pp_size,
+            tokenizer_manager.server_args.dp_size,
+            tokenizer_manager.server_args.enable_dp_attention,
+        ),
         "loads": loads,
     }
diff --git a/test/registered/unit/entrypoints/test_v1_loads_aggregate.py b/test/registered/unit/entrypoints/test_v1_loads_aggregate.py
index 66f059e88..ae31ae9b6 100644
--- a/test/registered/unit/entrypoints/test_v1_loads_aggregate.py
+++ b/test/registered/unit/entrypoints/test_v1_loads_aggregate.py
@@ -58,8 +58,21 @@ class _FakeTokenizerManager(TokenizerControlMixin):
 class _FakeHttpTokenizerManager:
     metrics_collector = None
 
-    def __init__(self, loads):
+    def __init__(
+        self,
+        loads,
+        tp_size=1,
+        dp_size=1,
+        pp_size=1,
+        enable_dp_attention=False,
+    ):
         self.loads = loads
+        self.server_args = SimpleNamespace(
+            tp_size=tp_size,
+            dp_size=dp_size,
+            pp_size=pp_size,
+            enable_dp_attention=enable_dp_attention,
+        )
 
     async def get_loads(self, include=None, dp_rank=None):
         results = []
@@ -94,16 +107,29 @@ class TestLoadsResponse(CustomTestCase):
 
 
 class TestLoadsAcceleratorField(CustomTestCase):
-    def test_accelerator_reported_in_json(self):
+    def test_accelerator_metadata_reported_in_json(self):
         """Guards the response contract: the JSON envelope carries an
-        "accelerator" field with the detected device name."""
-        manager = _FakeHttpTokenizerManager([LoadSnapshot(dp_rank=0)])
+        accelerator name and the accelerator count for each DP rank."""
+        manager = _FakeHttpTokenizerManager([LoadSnapshot(dp_rank=0)], tp_size=16)
 
         with mock.patch.object(
             v1_loads, "_accelerator_name", return_value="NVIDIA GB300"
         ):
             response = asyncio.run(get_loads(tokenizer_manager=manager))
             self.assertEqual(response["accelerator"], "NVIDIA GB300")
+            self.assertEqual(response["num_accelerators"], 16)
+
+    def test_dp_attention_reports_accelerators_per_dp_rank(self):
+        manager = _FakeHttpTokenizerManager(
+            [LoadSnapshot(dp_rank=rank) for rank in range(8)],
+            tp_size=8,
+            dp_size=8,
+            enable_dp_attention=True,
+        )
+
+        response = asyncio.run(get_loads(tokenizer_manager=manager))
+
+        self.assertEqual(response["num_accelerators"], 1)
 
 
 class TestGetLoads(CustomTestCase):