refactor model loader: initial refactor (#664)

python/sglang/srt/server_args.py

@@ -57,6 +57,7 @@ class ServerArgs:
     disable_disk_cache: bool = False
     attention_reduce_in_fp32: bool = False
     enable_p2p_check: bool = False
+    efficient_weight_load: bool = False
 
     # Distributed args
     nccl_init_addr: Optional[str] = None
@@ -327,6 +328,11 @@ class ServerArgs:
             action="store_true",
             help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
         )
+        parser.add_argument(
+            "--efficient-weight-load",
+            action="store_true",
+            help="Turn on memory efficient weight loading with quantization (quantize per layer during loading).",
+        )
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):