Emit consolidated Gemma4Audio op for gemma-4 audio configs

Follow the onnxruntime-extensions consolidation: both gemma4 audio configs
now use the single Gemma4Audio op with an explicit type attribute instead of
two separate op types.

* gemma4 / gemma4_text  -> Gemma4Audio type="log_mel"  (128-dim USM log-mel)
* gemma4_unified*       -> Gemma4Audio type="raw_frames" (raw 640-sample frames)

Updates the auto_export audio-config tests accordingly. 115 tests pass.

Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>

Author: justinchuby

src/mobius/integrations/ort_genai/auto_export.py

@@ -667,8 +667,8 @@ def _write_audio_processor_config(
         # Encoder-free unified model: each audio soft token is a raw chunk of the
         # 16 kHz waveform (audio_samples_per_token = audio_embed_dim, 640), not a
         # 128-dim log-mel frame. Reproduced natively by the ort-extensions
-        # ``Gemma4UnifiedAudioFrames`` op (pad to a whole number of frames,
-        # reshape to (num_tokens, 640)).
+        # ``Gemma4Audio`` op with ``type="raw_frames"`` (pad to a whole number of
+        # frames, reshape to (num_tokens, 640)).
         samples_per_token = getattr(audio, "hidden_size", None) or 640
         processor = {
             "feature_extraction": {
@@ -681,9 +681,10 @@ def _write_audio_processor_config(
                     },
                     {
                         "operation": {
-                            "name": "gemma4_unified_audio_frames",
-                            "type": "Gemma4UnifiedAudioFrames",
+                            "name": "gemma4_audio",
+                            "type": "Gemma4Audio",
                             "attrs": {
+                                "type": "raw_frames",
                                 "audio_samples_per_token": samples_per_token,
                                 "sampling_rate": 16000,
                                 "padding_value": 0.0,
@@ -695,7 +696,8 @@ def _write_audio_processor_config(
         }
         proc_filename = "audio_feature_extraction.json"
     elif model_type in _GEMMA4_MODEL_TYPES:
-        # Gemma4 USM-style 128-dim log-mel spectrogram.
+        # Gemma4 USM-style 128-dim log-mel spectrogram via the ort-extensions
+        # ``Gemma4Audio`` op with ``type="log_mel"``.
         # OrtxCreateSpeechFeatureExtractor requires the feature_extraction.sequence format.
         processor = {
             "feature_extraction": {
@@ -708,9 +710,10 @@ def _write_audio_processor_config(
                     },
                     {
                         "operation": {
-                            "name": "gemma4_log_mel",
-                            "type": "Gemma4LogMel",
+                            "name": "gemma4_audio",
+                            "type": "Gemma4Audio",
                             "attrs": {
+                                "type": "log_mel",
                                 "feature_size": 128,
                                 "sampling_rate": 16000,
                                 "frame_length_ms": 20.0,

src/mobius/integrations/ort_genai/auto_export_test.py

@@ -357,7 +357,8 @@ def test_audio_gemma4_unified_writes_raw_frames(self, tmp_path):
         seq = data["feature_extraction"]["sequence"]
         assert seq[0]["operation"]["type"] == "AudioDecoder"
         op = seq[1]["operation"]
-        assert op["type"] == "Gemma4UnifiedAudioFrames"
+        assert op["type"] == "Gemma4Audio"
+        assert op["attrs"]["type"] == "raw_frames"
         assert op["attrs"]["audio_samples_per_token"] == 640
 
     def test_audio_gemma4_writes_feature_extraction_json(self, tmp_path):
@@ -376,8 +377,9 @@ def test_audio_gemma4_writes_feature_extraction_json(self, tmp_path):
         seq = data["feature_extraction"]["sequence"]
         assert len(seq) == 2
         assert seq[0]["operation"]["type"] == "AudioDecoder"
-        assert seq[1]["operation"]["type"] == "Gemma4LogMel"
+        assert seq[1]["operation"]["type"] == "Gemma4Audio"
         attrs = seq[1]["operation"]["attrs"]
+        assert attrs["type"] == "log_mel"
         assert attrs["feature_size"] == 128
         assert attrs["sampling_rate"] == 16000
         assert attrs["frame_length_ms"] == 20.0  # noqa: RUF069
@@ -747,9 +749,10 @@ class FakeConfig:
         op0 = seq[0]["operation"]
         assert op0["type"] == "AudioDecoder"
 
-        # Second op: Gemma4LogMel with expected attrs
+        # Second op: Gemma4Audio (type=log_mel) with expected attrs
         op1 = seq[1]["operation"]
-        assert op1["type"] == "Gemma4LogMel"
+        assert op1["type"] == "Gemma4Audio"
+        assert op1["attrs"]["type"] == "log_mel"
         assert op1["attrs"]["feature_size"] == 128
         assert op1["attrs"]["sampling_rate"] == 16000
         assert op1["attrs"]["mel_floor"] == 0.001  # noqa: RUF069