Revert "Revert "Add simple CPU offloading support"" (#2253)

Co-authored-by: Jani Monoses <jani.monoses@gmail.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>

python/sglang/srt/model_executor/model_runner.py

@@ -61,6 +61,7 @@ from sglang.srt.utils import (
     is_hip,
     monkey_patch_vllm_model_config,
     monkey_patch_vllm_p2p_access_check,
+    set_cpu_offload_max_bytes,
 )
 
 logger = logging.getLogger(__name__)
@@ -145,6 +146,8 @@ class ModelRunner:
             }
         )
 
+        set_cpu_offload_max_bytes(int(server_args.cpu_offload_gb * 1024**3))
+
         # Init components
         min_per_gpu_memory = self.init_torch_distributed()
         self.sampler = Sampler()