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

2025-11-15 16:31:57 +08:00
parent 2a96e302cb
commit 56fc483073
2 changed files with 11 additions and 2 deletions
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/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,