Require compatible optimum-intel for OpenVINO conversion

olive/passes/openvino/optimum_intel.py

@@ -24,6 +24,20 @@
 _TEMP_DIR_LOCK = Lock()
 
 
+def _is_gptoss_quantization_config(quantization_config: Any) -> bool:
+    """Return whether the config is Optimum Intel's GPT-OSS-specific quantization config.
+
+    The private class was added after older supported optimum-intel releases. Treat it as
+    optional so those releases can continue using the normal quantization path.
+    """
+    try:
+        from optimum.intel.openvino.configuration import _GPTOSSQuantizationConfig
+    except ImportError:
+        return False
+
+    return isinstance(quantization_config, _GPTOSSQuantizationConfig)
+
+
 @contextmanager
 def _use_output_tempdir(output_model_path: str):
     with _TEMP_DIR_LOCK:
@@ -372,11 +386,7 @@ def _run_for_config(
     ) -> Union[OpenVINOModelHandler, CompositeModelHandler]:
         try:
             from optimum.exporters.openvino import main_export as export_optimum_intel
-            from optimum.intel.openvino.configuration import (
-                OVConfig,
-                _GPTOSSQuantizationConfig,
-                get_default_quantization_config,
-            )
+            from optimum.intel.openvino.configuration import OVConfig, get_default_quantization_config
             from optimum.intel.utils.import_utils import is_nncf_available
         except ImportError as e:
             raise ImportError("Please install Intel® optimum[openvino] to use OpenVINO Optimum Conversion") from e
@@ -497,8 +507,7 @@ def _run_for_config(
                         else:
                             if config.ov_quant_config.get("quantization_statistics_path", None) is not None:
                                 logger.warning(
-                                    "quantization_statistics_path is only applicable for weight-only"
-                                    " quantization. It will be ignored."
+                                    "quantization_statistics_path is only applicable for weight-only quantization. It will be ignored."
                                 )
                             quant_config = prep_q_config(config.ov_quant_config)
 
@@ -509,7 +518,7 @@ def _run_for_config(
         # quantization config
         quant_config = ov_config.quantization_config if ov_config else None
 
-        apply_main_quantize = quant_config and not isinstance(quant_config, _GPTOSSQuantizationConfig)
+        apply_main_quantize = quant_config and not _is_gptoss_quantization_config(quant_config)
 
         try:
             extra_args["ov_config"] = ov_config

test/passes/openvino/test_openvino_optimum_conversion.py

@@ -7,12 +7,32 @@
 import pytest
 
 from olive.passes.olive_pass import create_pass_from_dict
-from olive.passes.openvino.optimum_intel import OpenVINOOptimumConversion
+from olive.passes.openvino.optimum_intel import OpenVINOOptimumConversion, _is_gptoss_quantization_config
 from test.utils import get_hf_model, package_version_at_least
 
 pytestmark = pytest.mark.openvino
 
 
+def test_gptoss_quantization_config_is_optional(monkeypatch):
+    import optimum.intel.openvino.configuration as configuration
+
+    monkeypatch.delattr(configuration, "_GPTOSSQuantizationConfig", raising=False)
+
+    assert not _is_gptoss_quantization_config(object())
+
+
+def test_gptoss_quantization_config_is_detected_when_available(monkeypatch):
+    import optimum.intel.openvino.configuration as configuration
+
+    class FakeGPTOSSQuantizationConfig:
+        pass
+
+    monkeypatch.setattr(configuration, "_GPTOSSQuantizationConfig", FakeGPTOSSQuantizationConfig, raising=False)
+
+    assert _is_gptoss_quantization_config(FakeGPTOSSQuantizationConfig())
+    assert not _is_gptoss_quantization_config(object())
+
+
 @pytest.mark.skipif(
     not package_version_at_least("optimum", "2.1.0"),
     reason="Requires optimum >= 2.1.0",