Fix data-free pipeline quantization (#1477)

Author: nikita-savelyevv

optimum/intel/openvino/quantization.py

@@ -1335,7 +1335,7 @@ def _quantize_ovbasemodel(
                 default_config = OVWeightQuantizationConfig(bits=8, sym=True)
             else:
                 default_config = quantization_config
-        else:
+        elif not isinstance(quantization_config, OVPipelineQuantizationConfig):
             #
             # Hybrid/Full/Mixed quantization
             #
@@ -1397,7 +1397,7 @@ def _quantize_ovbasemodel(
                     raise NotImplementedError("Mixed precision quantization isn't supported for diffusers.")
 
                 default_config = quantization_config
-            elif not isinstance(quantization_config, OVPipelineQuantizationConfig):
+            else:
                 raise ValueError(f"Unsupported type of quantization config: {type(quantization_config)}")
 
         pipeline_quantization_config = (

tests/openvino/test_quantization.py

@@ -1487,6 +1487,14 @@ class OVPipelineQuantizationTest(unittest.TestCase):
     maxDiff = None
 
     PIPELINE_QUANTIZATION_SCOPE = [
+        (
+            OVModelForCausalLM,
+            "gpt2",
+            False,
+            dict(quantization_configs={"model": dict(bits=8, weight_only=True)}),
+            {"model": 0},
+            {"model": {"int8": 44}},
+        ),
         (
             OVModelForCausalLM,
             "llama",