[RL] support only do cpu backup on draft model (#13318)

python/sglang/srt/model_executor/model_runner.py

@@ -756,9 +756,12 @@ class ModelRunner:
         # Remove monkey_patch when linear.py quant remove dependencies with vllm
         monkey_patch_vllm_parallel_state()
 
+        enable_cpu_backup = self.server_args.enable_weights_cpu_backup or (
+            self.is_draft_worker and self.server_args.enable_draft_weights_cpu_backup
+        )
         with self.memory_saver_adapter.region(
             GPU_MEMORY_TYPE_WEIGHTS,
-            enable_cpu_backup=self.server_args.enable_weights_cpu_backup,
+            enable_cpu_backup=enable_cpu_backup,
         ):
             self.model = get_model(
                 model_config=self.model_config,

python/sglang/srt/server_args.py

@@ -511,6 +511,7 @@ class ServerArgs:
     delete_ckpt_after_loading: bool = False
     enable_memory_saver: bool = False
     enable_weights_cpu_backup: bool = False
+    enable_draft_weights_cpu_backup: bool = False
     allow_auto_truncate: bool = False
     enable_custom_logit_processor: bool = False
     flashinfer_mla_disable_ragged: bool = False
@@ -3409,7 +3410,12 @@ class ServerArgs:
         parser.add_argument(
             "--enable-weights-cpu-backup",
             action="store_true",
-            help="Save model weights to CPU memory during release_weights_occupation and resume_weights_occupation",
+            help="Save model weights (both main model and draft model, if any) to CPU memory during release_weights_occupation and resume_weights_occupation",
+        )
+        parser.add_argument(
+            "--enable-draft-weights-cpu-backup",
+            action="store_true",
+            help="Save draft model weights to CPU memory during release_weights_occupation and resume_weights_occupation",
         )
         parser.add_argument(
             "--allow-auto-truncate",