Add simple CPU offloading support. (#2081)

python/sglang/srt/server_args.py

@@ -62,6 +62,7 @@ class ServerArgs:
     max_prefill_tokens: int = 16384
     schedule_policy: str = "lpm"
     schedule_conservativeness: float = 1.0
+    cpu_offload_gb: int = 0
 
     # Other runtime options
     tp_size: int = 1
@@ -373,6 +374,13 @@ class ServerArgs:
             help="How conservative the schedule policy is. A larger value means more conservative scheduling. Use a larger value if you see requests being retracted frequently.",
         )
 
+        parser.add_argument(
+            "--cpu-offload-gb",
+            type=int,
+            default=ServerArgs.cpu_offload_gb,
+            help="How many GBs of RAM to reserve for CPU offloading",
+        )
+
         # Other runtime options
         parser.add_argument(
             "--tensor-parallel-size",