fix(loader): pass model_type to TasksManager for CTC ASR model class resolution

examples/recipes/jonatasgrosman_wav2vec2-large-xlsr-53-russian/cpu/cpu/automatic-speech-recognition_fp16_config.json

@@ -0,0 +1,44 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "input_values",
+        "dtype": "float32",
+        "shape": [
+          1,
+          16000
+        ],
+        "value_range": [
+          -1,
+          1
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "logits"
+      }
+    ]
+  },
+  "optim": {},
+  "quant": {
+    "precision": "fp16",
+    "op_types_to_quantize": null,
+    "nodes_to_exclude": null
+  },
+  "compile": null,
+  "loader": {
+    "task": "automatic-speech-recognition",
+    "model_class": "AutoModelForCTC",
+    "model_type": "wav2vec2"
+  }
+}

examples/recipes/jonatasgrosman_wav2vec2-large-xlsr-53-russian/cpu/cpu/automatic-speech-recognition_fp32_config.json

@@ -0,0 +1,40 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "input_values",
+        "dtype": "float32",
+        "shape": [
+          1,
+          16000
+        ],
+        "value_range": [
+          -1,
+          1
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "logits"
+      }
+    ]
+  },
+  "optim": {},
+  "quant": null,
+  "compile": null,
+  "loader": {
+    "task": "automatic-speech-recognition",
+    "model_class": "AutoModelForCTC",
+    "model_type": "wav2vec2"
+  }
+}

src/winml/modelkit/loader/resolution.py

@@ -644,7 +644,9 @@ def resolve_task(
         resolved = _get_custom_model_class(model_type_norm, opt_task)
         if resolved is None:
             try:
-                resolved = TasksManager.get_model_class_for_task(opt_task, framework="pt")
+                resolved = TasksManager.get_model_class_for_task(
+                    opt_task, framework="pt", model_type=model_type_norm or None
+                )
             except Exception:
                 resolved = _resolve_model_class_from_config(config)  # arch fallback
 

tests/unit/loader/test_resolve_task_and_model_class.py

@@ -117,3 +117,44 @@ def test_task_normalized_with_model_class(self):
         # Task is normalized (unlike the user-task path which preserves the original)
         assert r.task == "fill-mask"
         assert r.source == TaskSource.USER_CLASS
+
+
+class TestCTCModelClassResolution:
+    """CTC ASR models resolve to AutoModelForCTC, not AutoModelForSpeechSeq2Seq.
+
+    When automatic-speech-recognition maps to multiple candidate classes,
+    passing model_type enables Optimum to select the correct class per
+    architecture family.
+    """
+
+    @pytest.mark.parametrize(
+        "model_type,arch_name",
+        [
+            ("wav2vec2", "Wav2Vec2ForCTC"),
+            ("hubert", "HubertForCTC"),
+            ("data2vec-audio", "Data2VecAudioForCTC"),
+        ],
+    )
+    def test_ctc_architecture_resolves_to_auto_model_for_ctc(self, model_type, arch_name):
+        """ForCTC architectures resolve to AutoModelForCTC via model_type."""
+        config = MagicMock()
+        config.model_type = model_type
+        config.architectures = [arch_name]
+        config._name_or_path = ""
+
+        r = resolve_task(config)
+
+        assert r.task == "automatic-speech-recognition"
+        assert r.model_class.__name__ == "AutoModelForCTC"
+
+    def test_whisper_still_resolves_to_speech_seq2seq(self):
+        """Non-CTC ASR models (Whisper) still get AutoModelForSpeechSeq2Seq."""
+        config = MagicMock()
+        config.model_type = "whisper"
+        config.architectures = ["WhisperForConditionalGeneration"]
+        config._name_or_path = ""
+
+        r = resolve_task(config)
+
+        assert r.task == "automatic-speech-recognition"
+        assert r.model_class.__name__ == "AutoModelForSpeechSeq2Seq"