Fix transformers v4.35.0 compatibility (#471)

* fix trainer

* fix

* format

* fix version

Author: echarlaix

.github/workflows/test_inc.yml

@@ -30,7 +30,8 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install .[neural-compressor,ipex,diffusers,tests]
+        pip install .[neural-compressor,diffusers,tests]
+        pip install intel-extension-for-pytorch
     - name: Test with Pytest
       run: |
         pytest tests/neural_compressor/

optimum/intel/neural_compressor/quantization.py

@@ -15,6 +15,7 @@
 import copy
 import inspect
 import logging
+import warnings
 from enum import Enum
 from itertools import chain
 from pathlib import Path
@@ -30,16 +31,25 @@
 from neural_compressor.quantization import fit
 from torch.utils.data import DataLoader, RandomSampler
 from transformers import (
+    AutoModelForCausalLM,
+    AutoModelForMaskedLM,
+    AutoModelForMultipleChoice,
+    AutoModelForQuestionAnswering,
+    AutoModelForSeq2SeqLM,
+    AutoModelForSequenceClassification,
+    AutoModelForTokenClassification,
+    AutoModelForVision2Seq,
     DataCollator,
     PretrainedConfig,
     PreTrainedModel,
+    XLNetLMHeadModel,
     default_data_collator,
 )
 
 from optimum.exporters import TasksManager
 from optimum.exporters.onnx import OnnxConfig
 from optimum.onnxruntime import ORTModel
-from optimum.onnxruntime.modeling_decoder import ORTModelDecoder
+from optimum.onnxruntime.modeling_decoder import ORTModelForCausalLM
 from optimum.onnxruntime.modeling_seq2seq import ORTModelForConditionalGeneration
 from optimum.onnxruntime.utils import ONNX_DECODER_NAME
 from optimum.quantization_base import OptimumQuantizer
@@ -256,7 +266,7 @@ def quantize(
             if isinstance(self._original_model, ORTModelForConditionalGeneration):
                 raise RuntimeError("ORTModelForConditionalGeneration not supported for quantization")
 
-            if isinstance(self._original_model, ORTModelDecoder):
+            if isinstance(self._original_model, ORTModelForCausalLM):
                 model_or_path = self._original_model.onnx_paths
                 if len(model_or_path) > 1:
                     raise RuntimeError(
@@ -528,3 +538,49 @@ def _apply_quantization_from_config(q_config: Dict, model: torch.nn.Module) -> t
         q_model = convert(q_model, mapping=q_mapping, inplace=True)
 
     return q_model
+
+
+class IncQuantizedModel(INCModel):
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        warnings.warn(
+            f"The class `{cls.__name__}` has been depreciated and will be removed in optimum-intel v1.12, please use "
+            f"`{cls.__name__.replace('IncQuantized', 'INC')}` instead."
+        )
+        return super().from_pretrained(*args, **kwargs)
+
+
+class IncQuantizedModelForQuestionAnswering(IncQuantizedModel):
+    auto_model_class = AutoModelForQuestionAnswering
+
+
+class IncQuantizedModelForSequenceClassification(IncQuantizedModel):
+    auto_model_class = AutoModelForSequenceClassification
+
+
+class IncQuantizedModelForTokenClassification(IncQuantizedModel):
+    auto_model_class = AutoModelForTokenClassification
+
+
+class IncQuantizedModelForMultipleChoice(IncQuantizedModel):
+    auto_model_class = AutoModelForMultipleChoice
+
+
+class IncQuantizedModelForSeq2SeqLM(IncQuantizedModel):
+    auto_model_class = AutoModelForSeq2SeqLM
+
+
+class IncQuantizedModelForCausalLM(IncQuantizedModel):
+    auto_model_class = AutoModelForCausalLM
+
+
+class IncQuantizedModelForMaskedLM(IncQuantizedModel):
+    auto_model_class = AutoModelForMaskedLM
+
+
+class IncQuantizedModelForXLNetLM(IncQuantizedModel):
+    auto_model_class = XLNetLMHeadModel
+
+
+class IncQuantizedModelForVision2Seq(IncQuantizedModel):
+    auto_model_class = AutoModelForVision2Seq

optimum/intel/neural_compressor/trainer.py

@@ -13,7 +13,7 @@
 
 INSTALL_REQUIRE = [
     "optimum>=1.13.0",
-    "transformers>=4.20.0",
+    "transformers",
     "datasets>=1.4.0",
     "sentencepiece",
     "scipy",
@@ -41,8 +41,9 @@
         "neural-compressor>=2.2.0",
         "onnx",
         "onnxruntime<1.15.0",
+        "transformers>=4.33.0",
     ],
-    "openvino": ["openvino>=2023.1.0", "onnx", "onnxruntime"],
+    "openvino": ["openvino>=2023.1.0", "onnx", "onnxruntime", "transformers>=4.33.0"],
     "nncf": ["nncf>=2.6.0"],
     "ipex": ["transformers<4.32.0", "intel-extension-for-pytorch", "onnx"],
     "diffusers": ["diffusers"],

optimum/intel/openvino/trainer.py

@@ -19,6 +19,7 @@
 import unittest
 
 import torch
+from packaging.version import Version, parse
 from parameterized import parameterized
 from transformers import AutoTokenizer, pipeline, set_seed
 
@@ -39,6 +40,7 @@
     INCTrainer,
 )
 from optimum.intel.neural_compressor.utils import _HEAD_TO_AUTOMODELS, WEIGHTS_NAME
+from optimum.version import __version__ as _optimum_version
 
 
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
@@ -133,6 +135,7 @@ def test_pipeline(self, model_id, task):
 
         pipe(*inputs)
 
+    @unittest.skipIf(parse(_optimum_version) < Version("1.14.0"), "not supported, needs optimum>=v1.14.0")
     def test_compare_with_and_without_past_key_values(self):
         model_id = "echarlaix/tiny-random-gpt2-torchscript"
         tokenizer = AutoTokenizer.from_pretrained(model_id)

setup.py

@@ -54,7 +54,7 @@ def test_static_quantization(self, task, model_name, expected_quantized_matmuls)
             tokenizer.pad_token = tokenizer.eos_token
         quantizer = INCQuantizer.from_pretrained(model, task=task)
         calibration_dataset = _generate_dataset(quantizer, tokenizer, num_samples=num_samples)
-        save_onnx_model = True
+        save_onnx_model = False
         op_type_dict = (
             {"Embedding": {"weight": {"dtype": ["fp32"]}, "activation": {"dtype": ["fp32"]}}}
             if save_onnx_model