Config: do not sort positional arguments

thinc/config.py

@@ -1,5 +1,5 @@
 from typing import Union, Dict, Any, Optional, List, Tuple, Callable, Type
-from typing import Iterable, Sequence
+from typing import Iterable, Sequence, cast
 from types import GeneratorType
 from dataclasses import dataclass
 from configparser import ConfigParser, ExtendedInterpolation, MAX_INTERPOLATION_DEPTH
@@ -340,7 +340,10 @@ def _sort(
         account for subsections, which should always follow their parent.
         """
         sort_map = {section: i for i, section in enumerate(self.section_order)}
-        sort_key = lambda x: (sort_map.get(x[0].split(".")[0], len(sort_map)), x[0])
+        sort_key = lambda x: (
+            sort_map.get(x[0].split(".")[0], len(sort_map)),
+            _mask_positional_args(x[0]),
+        )
         return dict(sorted(data.items(), key=sort_key))
 
     def _set_overrides(self, config: "ConfigParser", overrides: Dict[str, Any]) -> None:
@@ -456,6 +459,20 @@ def from_disk(
         return self.from_str(text, interpolate=interpolate, overrides=overrides)
 
 
+def _mask_positional_args(name: str) -> List[Optional[str]]:
+    """Create a section name representation that masks names
+    of positional arguments to retain their order in sorts."""
+
+    stable_name = cast(List[Optional[str]], name.split("."))
+
+    # Remove names of sections that are a positional argument.
+    for i in range(1, len(stable_name)):
+        if stable_name[i - 1] == "*":
+            stable_name[i] = None
+
+    return stable_name
+
+
 def try_load_json(value: str) -> Any:
     """Load a JSON string if possible, otherwise default to original value."""
     try:

thinc/tests/test_config.py

@@ -1437,3 +1437,29 @@ def test_config_overrides(greeting, value, expected):
     assert "${vars.a}" in str_cfg
     cfg = Config().from_str(str_cfg, overrides=overrides)
     assert expected in str(cfg)
+
+
+def test_arg_order_is_preserved():
+    str_cfg = """
+    [model]
+
+    [model.chain]
+    @layers = "chain.v1"
+
+    [model.chain.*.hashembed]
+    @layers = "HashEmbed.v1"
+    nO = 8
+    nV = 8
+
+    [model.chain.*.expand_window]
+    @layers = "expand_window.v1"
+    window_size = 1
+    """
+
+    cfg = Config().from_str(str_cfg)
+    resolved = my_registry.resolve(cfg)
+    model = resolved["model"]["chain"]
+
+    # Fails when arguments are sorted, because expand_window
+    # is sorted before hashembed.
+    assert model.name == "hashembed>>expand_window"