Increase default 4-bit compression ratio from 0.8 to 1.0 (#805)

* Increase default 4-bit ratio from 0.8 to 1.0

* Style

* Fix test

Author: nikita-savelyevv

optimum/commands/export/openvino.py

@@ -221,7 +221,7 @@ def parse_args(parser: "ArgumentParser"):
 
     def run(self):
         from ...exporters.openvino.__main__ import infer_task, main_export, maybe_convert_tokenizers
-        from ...intel.openvino.configuration import _DEFAULT_4BIT_CONFIGS, OVConfig
+        from ...intel.openvino.configuration import _DEFAULT_4BIT_CONFIG, _DEFAULT_4BIT_CONFIGS, OVConfig
 
         def _get_default_int4_config(model_id_or_path, library_name):
             if model_id_or_path in _DEFAULT_4BIT_CONFIGS:
@@ -233,13 +233,7 @@ def _get_default_int4_config(model_id_or_path, library_name):
                 if original_model_name in _DEFAULT_4BIT_CONFIGS:
                     return _DEFAULT_4BIT_CONFIGS[original_model_name]
 
-            return {
-                "bits": 4,
-                "ratio": 0.8,
-                "sym": False,
-                "group_size": None,
-                "all_layers": None,
-            }
+            return _DEFAULT_4BIT_CONFIG
 
         library_name = TasksManager.infer_library_from_model(self.args.model, library_name=self.args.library)
         if library_name == "sentence_transformers" and self.args.library is None:

optimum/intel/openvino/configuration.py

@@ -104,6 +104,14 @@
     },
 }
 
+_DEFAULT_4BIT_CONFIG = {
+    "bits": 4,
+    "ratio": 1.0,
+    "sym": False,
+    "group_size": 128,
+    "all_layers": None,
+}
+
 
 class OVQuantizationMethod(str, Enum):
     DEFAULT = "default"

optimum/intel/openvino/modeling_decoder.py

@@ -40,7 +40,13 @@
 from ...exporters.openvino.stateful import model_has_state
 from ..utils.import_utils import is_nncf_available, is_transformers_version
 from ..utils.modeling_utils import MULTI_QUERY_ATTN_MODELS
-from .configuration import _DEFAULT_4BIT_CONFIGS, OVConfig, OVWeightQuantizationConfig, _check_default_4bit_configs
+from .configuration import (
+    _DEFAULT_4BIT_CONFIG,
+    _DEFAULT_4BIT_CONFIGS,
+    OVConfig,
+    OVWeightQuantizationConfig,
+    _check_default_4bit_configs,
+)
 from .modeling import _TOKENIZER_FOR_DOC, INPUTS_DOCSTRING, MODEL_START_DOCSTRING, OVModel
 from .utils import ONNX_WEIGHTS_NAME, OV_TO_NP_TYPE, OV_XML_FILE_NAME, STR_TO_OV_TYPE
 
@@ -775,7 +781,7 @@ def _from_pretrained(
             init_cls = cls
 
         if isinstance(quantization_config, dict) and quantization_config == {"bits": 4}:
-            quantization_config = _DEFAULT_4BIT_CONFIGS.get(config.name_or_path, quantization_config)
+            quantization_config = _DEFAULT_4BIT_CONFIGS.get(config.name_or_path, _DEFAULT_4BIT_CONFIG)
         quantization_config = cls._prepare_weight_quantization_config(quantization_config, load_in_8bit)
 
         enable_compilation = kwargs.pop("compile", True) and not quantization_config

tests/openvino/test_exporters_cli.py

@@ -86,10 +86,10 @@ class OVCLIExportTestCase(unittest.TestCase):
     )
 
     TEST_4BIT_CONFIGURATONS = [
-        ("text-generation-with-past", "opt125m", "int4_sym_g128", 62, 86),
-        ("text-generation-with-past", "opt125m", "int4_asym_g128", 62, 86),
-        ("text-generation-with-past", "opt125m", "int4_sym_g64", 62, 86),
-        ("text-generation-with-past", "opt125m", "int4_asym_g64", 62, 86),
+        ("text-generation-with-past", "opt125m", "int4_sym_g128", 4, 144),
+        ("text-generation-with-past", "opt125m", "int4_asym_g128", 4, 144),
+        ("text-generation-with-past", "opt125m", "int4_sym_g64", 4, 144),
+        ("text-generation-with-past", "opt125m", "int4_asym_g64", 4, 144),
         ("text-generation-with-past", "llama_awq", "int4 --ratio 1.0 --sym --group-size 16 --all-layers", 0, 32),
         (
             "text-generation-with-past",