fix(args): skip Megatron validate_args for external SGLang server

miles/utils/arguments.py

@@ -21,6 +21,17 @@
 logger = logging.getLogger(__name__)
 
 
+def is_external_model_server(args) -> bool:
+    """Return whether rollout runs against an external SGLang model server.
+
+    In this mode ``sglang_router_ip`` points at a long-lived remote server and no
+    local rollout GPUs are allocated, so the framework should skip Megatron argument
+    validation (which reads CUDA device properties) to allow evals to run on
+    CPU-only pods with no NVIDIA driver.
+    """
+    return getattr(args, "sglang_router_ip", None) is not None and args.rollout_num_gpus == 0
+
+
 def reset_arg(parser, name, **kwargs):
     """
     Reset the default value of a Megatron argument.
@@ -2170,7 +2181,7 @@ def parse_args(add_custom_arguments=None):
 
     miles_validate_args(args)
 
-    if backend == "megatron":
+    if backend == "megatron" and not is_external_model_server(args):
         megatron_validate_args(args)
 
         # always use varlen