[Router] Report accelerator count in /v1/loads (#33548)

Co-authored-by: Yinghai Lu <yinghai@meta.com>

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,
     }