Fix default int8 quantization for CLI (#592)

Author: echarlaix

optimum/commands/export/openvino.py

@@ -157,13 +157,12 @@ def run(self):
             )
             self.args.weight_format = "int8"
 
-        weight_format = self.args.weight_format or "fp32"
-
-        ov_config = None
-        if weight_format in {"fp16", "fp32"}:
-            ov_config = OVConfig(dtype=weight_format)
+        if self.args.weight_format is None:
+            ov_config = None
+        elif self.args.weight_format in {"fp16", "fp32"}:
+            ov_config = OVConfig(dtype=self.args.weight_format)
         else:
-            is_int8 = weight_format == "int8"
+            is_int8 = self.args.weight_format == "int8"
 
             # For int4 quantization if not parameter is provided, then use the default config if exist
             if (
@@ -182,12 +181,12 @@ def run(self):
                     "group_size": -1 if is_int8 else self.args.group_size,
                 }
 
-            if weight_format in {"int4_sym_g128", "int4_asym_g128", "int4_sym_g64", "int4_asym_g64"}:
+            if self.args.weight_format in {"int4_sym_g128", "int4_asym_g128", "int4_sym_g64", "int4_asym_g64"}:
                 logger.warning(
-                    f"--weight-format {weight_format} is deprecated, possible choices are fp32, fp16, int8, int4"
+                    f"--weight-format {self.args.weight_format} is deprecated, possible choices are fp32, fp16, int8, int4"
                 )
-                quantization_config["sym"] = "asym" not in weight_format
-                quantization_config["group_size"] = 128 if "128" in weight_format else 64
+                quantization_config["sym"] = "asym" not in self.args.weight_format
+                quantization_config["group_size"] = 128 if "128" in self.args.weight_format else 64
             ov_config = OVConfig(quantization_config=quantization_config)
 
         # TODO : add input shapes

optimum/exporters/openvino/__main__.py

@@ -21,26 +21,13 @@
 
 from optimum.exporters import TasksManager
 from optimum.exporters.onnx.base import OnnxConfig
+from optimum.exporters.onnx.constants import SDPA_ARCHS_ONNX_EXPORT_NOT_SUPPORTED
 from optimum.utils.save_utils import maybe_load_preprocessors
 
-from ...intel.utils.import_utils import (
-    is_openvino_tokenizers_available,
-    is_optimum_version,
-    is_transformers_version,
-)
+from ...intel.utils.import_utils import is_openvino_tokenizers_available, is_transformers_version
 from .convert import export_from_model, export_tokenizer
 
 
-if is_optimum_version(">=", "1.16.0"):
-    from optimum.exporters.onnx.constants import SDPA_ARCHS_ONNX_EXPORT_NOT_SUPPORTED
-else:
-    # Copied from https://github.com/huggingface/optimum/blob/main/optimum/exporters/onnx/constants.py
-    SDPA_ARCHS_ONNX_EXPORT_NOT_SUPPORTED = [
-        "bart",
-        "whisper",
-    ]
-
-
 if TYPE_CHECKING:
     from optimum.intel.openvino.configuration import OVConfig