fix(cli): stringify list values from a YAML/JSON config before building argv

swift/cli/main.py

@@ -61,7 +61,7 @@ def parse_yaml_args(argv):
     for k, v in config.items():
         config_argv.append(f'--{k}')
         if isinstance(v, list):
-            config_argv += v
+            config_argv += [str(i) for i in v]
         else:
             if isinstance(v, dict):
                 v = json.dumps(v, ensure_ascii=False)

tests/utils/test_cli_config_args.py

@@ -0,0 +1,61 @@
+import os
+import tempfile
+import unittest
+
+from swift.cli.main import parse_yaml_args
+
+
+class TestYamlConfigArgv(unittest.TestCase):
+    """`swift sft config.yaml` (and the JSON form) rewrites argv in place, and the launcher then
+    prints and execs it as one command line, so every value has to arrive as a string."""
+
+    def setUp(self):
+        self._dir = tempfile.TemporaryDirectory()
+        self.addCleanup(self._dir.cleanup)
+        # parse_yaml_args records the config path for the entry point that runs after it
+        self._saved_config = os.environ.pop('SWIFT_CONFIG_FILE', None)
+        self.addCleanup(self._restore_config)
+
+    def _restore_config(self):
+        os.environ.pop('SWIFT_CONFIG_FILE', None)
+        if self._saved_config is not None:
+            os.environ['SWIFT_CONFIG_FILE'] = self._saved_config
+
+    def parse(self, name, content):
+        path = os.path.join(self._dir.name, name)
+        with open(path, 'w', encoding='utf-8') as f:
+            f.write(content)
+        argv = [path]
+        parse_yaml_args(argv)
+        return argv
+
+    def assert_launchable(self, argv):
+        try:
+            ' '.join(argv)
+        except TypeError as exc:
+            self.fail(f'the launcher cannot build a command line from {argv!r}: {exc}')
+
+    def test_float_list_is_passed_as_strings(self):
+        # e.g. `interleave_prob: [0.5, 0.5]`, declared Optional[List[float]] in DataArguments
+        argv = self.parse('sft.yaml', 'dataset:\n- a\n- b\ninterleave_prob: [0.5, 0.5]\n')
+        self.assert_launchable(argv)
+        self.assertEqual(argv[-3:], ['--interleave_prob', '0.5', '0.5'])
+
+    def test_int_list_is_passed_as_strings(self):
+        # e.g. `data_range: [0, 2]`, declared List[int] in SamplingArguments
+        argv = self.parse('sample.json', '{"num_samples": 10, "data_range": [0, 2]}')
+        self.assert_launchable(argv)
+        self.assertEqual(argv, ['--num_samples', '10', '--data_range', '0', '2'])
+
+    def test_string_list_is_unchanged(self):
+        argv = self.parse('datasets.yaml', 'dataset:\n- a#100\n- b#100\n')
+        self.assertEqual(argv, ['--dataset', 'a#100', 'b#100'])
+
+    def test_dict_value_is_still_serialized(self):
+        argv = self.parse('engine.yaml', 'model: m\nvllm_engine_kwargs:\n  gpu_memory_utilization: 0.8\n')
+        self.assert_launchable(argv)
+        self.assertEqual(argv, ['--model', 'm', '--vllm_engine_kwargs', '{"gpu_memory_utilization": 0.8}'])
+
+
+if __name__ == '__main__':
+    unittest.main()