Support min load besides random routing key assignment policy in ManualPolicy (#16767)

sgl-model-gateway/bindings/python/src/sglang_router/router_args.py

@@ -36,6 +36,7 @@ class RouterArgs:
     eviction_interval_secs: int = 60
     max_tree_size: int = 2**26
     max_idle_secs: int = 4 * 3600
+    assignment_mode: str = "random"
     max_payload_size: int = 512 * 1024 * 1024  # 512MB default for large batches
     bucket_adjust_interval_secs: int = 5
     dp_aware: bool = False
@@ -310,6 +311,13 @@ class RouterArgs:
             default=RouterArgs.max_idle_secs,
             help="Maximum idle time in seconds before eviction (for manual policy)",
         )
+        routing_group.add_argument(
+            f"--{prefix}assignment-mode",
+            type=str,
+            default=RouterArgs.assignment_mode,
+            choices=["random", "min_load"],
+            help="Mode for assigning new routing keys in manual policy: random (default), min_load (worker with fewest requests)",
+        )
         routing_group.add_argument(
             f"--{prefix}max-payload-size",
             type=int,