[Feature] Support loading weights from ckpt engine worker (#11755)

Signed-off-by: Yang Kaiyong <yangkaiyong.yky@antgroup.com>
Signed-off-by: Cruz Zhao <CruzZhao@linux.alibaba.com>
Signed-off-by: Xuchun Shang <xuchun.shang@gmail.com>
Co-authored-by: Yang Kaiyong <yangkaiyong.yky@antgroup.com>
Co-authored-by: Cruz Zhao <CruzZhao@linux.alibaba.com>
Co-authored-by: Xuchun Shang <xuchun.shang@gmail.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>

python/sglang/srt/server_args.py

@@ -208,6 +208,7 @@ class ServerArgs:
     skip_server_warmup: bool = False
     warmups: Optional[str] = None
     nccl_port: Optional[int] = None
+    checkpoint_engine_wait_weights_before_ready: bool = False
 
     # Quantization and data type
     dtype: str = "auto"
@@ -1704,6 +1705,12 @@ class ServerArgs:
             default=ServerArgs.nccl_port,
             help="The port for NCCL distributed environment setup. Defaults to a random port.",
         )
+        parser.add_argument(
+            "--checkpoint-engine-wait-weights-before-ready",
+            action="store_true",
+            help="If set, the server will wait for initial weights to be loaded via checkpoint-engine or other update methods "
+            "before serving inference requests.",
+        )
 
         # Quantization and data type
         parser.add_argument(