fix: apply config-file overrides before argument validation

miles/utils/arguments.py

@@ -2289,7 +2289,9 @@ def add_user_provided_function_arguments(parser):
             "--custom-config-path",
             type=str,
             default=None,
-            help="Path to the YAML config for custom function arguments.",
+            help="Path to the YAML config for custom function arguments. Entries are applied onto "
+            "the parsed args before validation, so overriding a core argument behaves like "
+            "passing it on the command line.",
         )
         reset_arg(parser, "--padded-vocab-size", type=int, default=None)
 
@@ -2431,6 +2433,11 @@ def _resolve_ft_components(args: argparse.Namespace) -> list[str]:
 
 
 def miles_validate_args(args):
+    # Config-file overrides must land before any validation or derivation below,
+    # so that derived flags (e.g. use_critic from advantage_estimator) and the
+    # worker topology are computed from the overridden values.
+    _apply_custom_config_overrides(args)
+
     validate_dashboard_args(args)
 
     args.ft_components = _resolve_ft_components(args)
@@ -2900,14 +2907,6 @@ def miles_validate_args(args):
     if args.use_rollout_routing_replay:
         args.use_routing_replay = True
 
-    if args.custom_config_path:
-        with open(args.custom_config_path) as f:
-            data = yaml.safe_load(f) or {}
-        for k, v in data.items():
-            if hasattr(args, k):
-                logger.info(f"Warning: Argument {k} is already set to {getattr(args, k)}, will override with {v}.")
-            setattr(args, k, v)
-
     if args.use_rollout_indexer_replay:
         args.use_indexer_replay = True
         assert args.context_parallel_size == 1, "indexer replay does not support context parallelism yet"
@@ -2970,6 +2969,25 @@ def validate_async_off_policy_correction(args) -> None:
     )
 
 
+def _apply_custom_config_overrides(args) -> None:
+    """Apply --custom-config-path YAML entries onto the parsed args.
+
+    Runs at the very start of `miles_validate_args`, so an overridden argument
+    behaves as if it had been passed on the command line: it participates in
+    every subsequent validation and derivation instead of silently bypassing
+    them. Tests call `miles_validate_args` with partial namespaces, so a
+    missing attribute is treated as "no config file".
+    """
+    if not getattr(args, "custom_config_path", None):
+        return
+    with open(args.custom_config_path) as f:
+        data = yaml.safe_load(f) or {}
+    for k, v in data.items():
+        if hasattr(args, k):
+            logger.info(f"Warning: Argument {k} is already set to {getattr(args, k)}, will override with {v}.")
+        setattr(args, k, v)
+
+
 def _maybe_apply_dumper_overrides(args) -> None:
     if not args.dumper_enable:
         return

tests/fast/utils/test_arguments.py

@@ -9,6 +9,7 @@
 from miles.backends.sglang_utils.arguments import add_sglang_arguments
 from miles.backends.sglang_utils.arguments import validate_args as validate_sglang_args
 from miles.utils.arguments import (
+    _apply_custom_config_overrides,
     _maybe_apply_dumper_overrides,
     _resolve_ft_components,
     get_miles_extra_args_provider,
@@ -496,3 +497,30 @@ def test_ppo_with_any_correction_passes(self, flag):
 
     def test_non_ppo_estimators_are_unaffected(self):
         validate_async_off_policy_correction(_make_async_ppo_args(use_critic=False))
+
+
+class TestApplyCustomConfigOverrides:
+    def test_no_config_path_is_a_noop(self):
+        args = SimpleNamespace(custom_config_path=None, advantage_estimator="grpo")
+        _apply_custom_config_overrides(args)
+        assert args.advantage_estimator == "grpo"
+
+    def test_missing_attribute_is_a_noop(self):
+        args = SimpleNamespace(advantage_estimator="grpo")
+        _apply_custom_config_overrides(args)
+        assert args.advantage_estimator == "grpo"
+
+    def test_empty_config_is_a_noop(self, tmp_path):
+        config = tmp_path / "overrides.yaml"
+        config.write_text("")
+        args = SimpleNamespace(custom_config_path=str(config), advantage_estimator="grpo")
+        _apply_custom_config_overrides(args)
+        assert args.advantage_estimator == "grpo"
+
+    def test_overrides_existing_and_adds_new_keys(self, tmp_path):
+        config = tmp_path / "overrides.yaml"
+        config.write_text("advantage_estimator: ppo\nmy_custom_function_arg: 7\n")
+        args = SimpleNamespace(custom_config_path=str(config), advantage_estimator="grpo")
+        _apply_custom_config_overrides(args)
+        assert args.advantage_estimator == "ppo"
+        assert args.my_custom_function_arg == 7