fix(model): rewrite seq_cls architectures to *ForSequenceClassification

swift/model/patcher.py

@@ -249,6 +249,29 @@ def transformers_seq_cls_forward(self, *args, origin_forward, padding_side=None,
     )
 
 
+def _seq_cls_architectures(arch_list):
+    """Rewrite architectures to the matching ``*ForSequenceClassification`` class.
+
+    Used by :func:`_patch_sequence_classification` so that the on-disk
+    ``config.json`` produced by ``PreTrainedModel.save_pretrained`` references
+    the seq_cls architecture (e.g. ``Qwen3VLForSequenceClassification``) instead
+    of the generation architecture that the model class is actually an instance
+    of. Without this, downstream vLLM deployment fails because the checkpoint
+    advertises ``Qwen3VLForConditionalGeneration`` while shipping a
+    ``score`` head and ``num_labels`` / ``problem_type`` fields — see #9704.
+    """
+    if not arch_list:
+        return arch_list
+    res = []
+    for arch in arch_list:
+        if arch.endswith('ForConditionalGeneration'):
+            arch = arch[:-len('ForConditionalGeneration')] + 'ForSequenceClassification'
+        elif arch.endswith('ForCausalLM'):
+            arch = arch[:-len('ForCausalLM')] + 'ForSequenceClassification'
+        res.append(arch)
+    return res
+
+
 def _patch_sequence_classification(model, model_meta):
     hidden_size = HfConfigFactory.get_config_attr(model.config, 'hidden_size')
     initializer_range = HfConfigFactory.get_config_attr(model.config, 'initializer_range')
@@ -274,6 +297,12 @@ def new_forward(self, *args, **kwargs):
 
     lm_head_model.forward = MethodType(new_forward, lm_head_model)
 
+    # Align the on-disk `architectures` with the task. PreTrainedModel.save_pretrained
+    # writes `model.__class__.__name__` for `architectures`, but the seq_cls patcher
+    # monkey-patches a `score` head onto the generation class without swapping it,
+    # so the saved checkpoint would otherwise advertise the wrong class (see #9704).
+    model.config.architectures = _seq_cls_architectures(getattr(model.config, 'architectures', None))
+
 
 @contextmanager
 def patch_automodel_for_sequence_classification(model_info=None,

tests/general/test_model.py

@@ -48,6 +48,53 @@ def test_registration(self):
         self.assertIn(TemplateType.molmo2, TEMPLATE_MAPPING)
 
 
+class TestSeqClsArchitecturesRewrite(unittest.TestCase):
+    """Regression coverage for #9704.
+
+    The seq_cls / reranker patcher monkey-patches a ``score`` head onto the
+    generation model class without swapping the class itself, so
+    ``PreTrainedModel.save_pretrained`` would otherwise write the wrong
+    ``architectures`` to ``config.json`` and break downstream vLLM deployment.
+    """
+
+    def test_seq_cls_architectures_rewrite(self):
+        from swift.model.patcher import _seq_cls_architectures
+
+        # Generation -> SequenceClassification
+        self.assertEqual(_seq_cls_architectures(['Qwen2ForCausalLM']), ['Qwen2ForSequenceClassification'])
+        self.assertEqual(
+            _seq_cls_architectures(['Qwen3VLForConditionalGeneration']),
+            ['Qwen3VLForSequenceClassification'])
+        self.assertEqual(_seq_cls_architectures(['LlamaForCausalLM']), ['LlamaForSequenceClassification'])
+        self.assertEqual(
+            _seq_cls_architectures(['Qwen2VLForConditionalGeneration']),
+            ['Qwen2VLForSequenceClassification'])
+
+        # Already-seq_cls class is preserved (idempotent).
+        self.assertEqual(
+            _seq_cls_architectures(['BertForSequenceClassification']),
+            ['BertForSequenceClassification'])
+
+        # Multi-arch list: only the matching suffix is rewritten.
+        self.assertEqual(
+            _seq_cls_architectures(['FooForCausalLM', 'BarForConditionalGeneration']),
+            ['FooForSequenceClassification', 'BarForSequenceClassification'])
+
+        # Empty / None inputs are returned unchanged.
+        self.assertEqual(_seq_cls_architectures([]), [])
+        self.assertIsNone(_seq_cls_architectures(None))
+
+    def test_seq_cls_architectures_unknown_suffix(self):
+        """Unknown suffixes (custom architectures) are left alone — we don't
+        silently invent a class name that may not exist in transformers."""
+        from swift.model.patcher import _seq_cls_architectures
+
+        self.assertEqual(_seq_cls_architectures(['MyCustomLMHeadModel']), ['MyCustomLMHeadModel'])
+        self.assertEqual(
+            _seq_cls_architectures(['MyCustomLMHeadModel', 'LlamaForCausalLM']),
+            ['MyCustomLMHeadModel', 'LlamaForSequenceClassification'])
+
+
 if __name__ == '__main__':
     test_qwen2()
     # test_modelscope_hub()