Merge pull request #673 from nikita-savelyevv/nncf-210-update

Update NNCF to 2.10. Enable AWQ algorithm.

Author: AlexKoff88

optimum/intel/openvino/quantization.py

@@ -37,6 +37,7 @@
 from transformers import AutoTokenizer, DataCollator, PreTrainedModel, default_data_collator
 from transformers.pytorch_utils import Conv1D
 from transformers.utils import is_accelerate_available
+from transformers.utils.quantization_config import QuantizationMethod
 
 from optimum.exporters.onnx.convert import check_dummy_inputs_are_allowed
 from optimum.exporters.tasks import TasksManager
@@ -670,10 +671,10 @@ def _weight_only_quantization(
         group_size=config.group_size,
         all_layers=config.all_layers,
         sensitivity_metric=sensitivity_metric,
-        # awq=config.quant_method == QuantizationMethod.AWQ,    # TODO : enable from nncf v2.9.0
+        awq=config.quant_method == QuantizationMethod.AWQ or None,
         ignored_scope=config.get_ignored_scope_instance(),
         dataset=dataset,
-        # subset_size=config.num_samples if config.num_samples else 128,    # TODO : enable from nncf v2.9.0
+        subset_size=config.num_samples if config.num_samples else 128,
     )
 
 

setup.py

@@ -60,8 +60,8 @@
 
 EXTRAS_REQUIRE = {
     "neural-compressor": ["neural-compressor>=2.2.0", "onnxruntime<1.15.0", "accelerate"],
-    "openvino": ["openvino>=2023.3", "nncf>=2.8.1", "openvino-tokenizers[transformers]"],
-    "nncf": ["nncf>=2.8.1"],
+    "openvino": ["openvino>=2023.3", "nncf>=2.10.0", "openvino-tokenizers[transformers]"],
+    "nncf": ["nncf>=2.10.0"],
     "ipex": ["intel-extension-for-pytorch", "transformers>=4.36.0,<4.39.0"],
     "diffusers": ["diffusers"],
     "quality": QUALITY_REQUIRE,

tests/openvino/test_quantization.py

@@ -221,17 +221,17 @@ class OVWeightCompressionTest(unittest.TestCase):
         ),
         (
             OVModelForCausalLM,
-            "opt",
+            "llama_awq",
             dict(
                 bits=4,
                 sym=True,
-                group_size=-1,
+                group_size=16,
                 ratio=0.8,
                 sensitivity_metric="mean_activation_magnitude",
                 dataset="ptb",
                 quant_method=QuantizationMethod.AWQ,
             ),
-            14,
+            16,
         ),
     )
 
@@ -452,6 +452,10 @@ def test_ovmodel_4bit_auto_compression_with_config(
         with tempfile.TemporaryDirectory() as tmp_dir:
             quantization_config = OVWeightQuantizationConfig.from_dict(quantization_config)
             model = model_cls.from_pretrained(model_id, export=True, quantization_config=quantization_config)
+            if quantization_config.quant_method == QuantizationMethod.AWQ:
+                # TODO: Check that AWQ was actually applied
+                pass
+
             tokenizer = AutoTokenizer.from_pretrained(model_id)
             if tokenizer.pad_token is None:
                 tokenizer.pad_token = tokenizer.eos_token
@@ -548,6 +552,8 @@ def test_ovmodel_load_large_model_with_additional_quantization_config(self):
                             "sensitivity_metric": None,
                             "dataset": None,
                             "ignored_scope": nncf.IgnoredScope(),
+                            "awq": None,
+                            "subset_size": 128,
                         }
                         compress_weights_patch.assert_called_with(unittest.mock.ANY, **compression_params)
 

tests/openvino/utils_tests.py

@@ -57,6 +57,7 @@
     "levit": "hf-internal-testing/tiny-random-LevitModel",
     "longt5": "hf-internal-testing/tiny-random-longt5",
     "llama": "fxmarty/tiny-llama-fast-tokenizer",
+    "llama_awq": "HuggingFaceH4/tiny-random-LlamaForCausalLM",
     "llama_gptq": "hf-internal-testing/TinyLlama-1.1B-Chat-v0.3-GPTQ",
     "m2m_100": "hf-internal-testing/tiny-random-m2m_100",
     "opt": "hf-internal-testing/tiny-random-OPTModel",