Add INC WoQ and remove ITREX dependency (huggingface#880)

* add inc woq and remove itrex dependency Signed-off-by: changwangss <[email protected]> * Update optimum/intel/neural_compressor/modeling_base.py Co-authored-by: Ella Charlaix <[email protected]> * Update optimum/intel/neural_compressor/modeling_base.py Co-authored-by: Ella Charlaix <[email protected]> * Update optimum/intel/neural_compressor/modeling_base.py Co-authored-by: Ella Charlaix <[email protected]> * Update optimum/intel/neural_compressor/modeling_base.py Co-authored-by: Ella Charlaix <[email protected]> * fix code according comment Signed-off-by: changwangss <[email protected]> * add logger setting Signed-off-by: changwangss <[email protected]> * improve ut Signed-off-by: changwangss <[email protected]> * move woq quantization to quantization.py Signed-off-by: changwangss <[email protected]> * Update examples/neural_compressor/language-modeling/run_clm.py Co-authored-by: Ilyas Moutawwakil <[email protected]> * Update examples/neural_compressor/language-modeling/run_clm.py Co-authored-by: Ilyas Moutawwakil <[email protected]> * remove dependency Signed-off-by: changwangss <[email protected]> * Update examples/neural_compressor/language-modeling/run_clm.py * add woq saving and loading ut and logger info Signed-off-by: changwangss <[email protected]> * set transformers version limit Signed-off-by: changwangss <[email protected]> * fix installation neural_compressor[pt] Signed-off-by: changwangss <[email protected]> * improve ut Signed-off-by: changwangss <[email protected]> * refactoring * Refactor * revert * fix datasets loading issue Signed-off-by: changwangss <[email protected]> * fix --------- Signed-off-by: changwangss <[email protected]> Co-authored-by: Ella Charlaix <[email protected]> Co-authored-by: Ilyas Moutawwakil <[email protected]> Co-authored-by: Ella Charlaix <[email protected]>
jiqing-feng · Sep 9, 2024 · 8a015a6 · 8a015a6
1 parent 4dc4d57
commit 8a015a6
Show file tree

Hide file tree

Showing 11 changed files with 251 additions and 277 deletions.
diff --git a/.github/workflows/test_inc.yml b/.github/workflows/test_inc.yml
@@ -36,16 +36,15 @@ jobs:
         pip install cmake
         pip install py-cpuinfo
         pip install torch==2.3.0 torchaudio==2.3.0 torchvision==0.18 --index-url https://download.pytorch.org/whl/cpu
+        pip install intel-extension-for-pytorch==2.3.0
+        pip install datasets==2.19.0
         pip install .[neural-compressor,diffusers,tests]
-        pip install intel-extension-for-transformers
         pip install peft
 
     - name: Test with Pytest
       run: |
         pytest tests/neural_compressor/ --ignore tests/neural_compressor/test_ipex.py --durations=0
     - name: Test IPEX
       run: |
-        pip uninstall -y intel-extension-for-transformers
-        pip install intel-extension-for-pytorch==2.3.0
         pytest tests/neural_compressor/test_ipex.py
 
diff --git a/examples/neural_compressor/language-modeling/README.md b/examples/neural_compressor/language-modeling/README.md
@@ -97,4 +97,4 @@ respectively `dynamic`, `static`, `weight_only` or `aware_training`.
 
 The flag `--verify_loading` can be passed along to verify that the resulting quantized model can be loaded correctly.
 
-> **_Note:_** `weight_only` quantization_approach requires `neural-compressor` >= 2.3 and `intel-extension-for-transformers` >= 1.3.
+> **_Note:_** `weight_only` quantization_approach requires `neural-compressor` > 3.0.
diff --git a/examples/neural_compressor/language-modeling/requirements.txt b/examples/neural_compressor/language-modeling/requirements.txt
@@ -3,5 +3,4 @@ torch >= 1.9
 datasets >= 1.8.0
 sentencepiece != 0.1.92
 protobuf
-intel-extension-for-transformers >= 1.3
 peft
diff --git a/examples/neural_compressor/language-modeling/run_clm.py b/examples/neural_compressor/language-modeling/run_clm.py
@@ -39,6 +39,7 @@
     QuantizationAwareTrainingConfig,
     WeightPruningConfig,
 )
+from neural_compressor.transformers import GPTQConfig, RtnConfig
 from transformers import (
     CONFIG_MAPPING,
     MODEL_FOR_CAUSAL_LM_MAPPING,
@@ -57,12 +58,8 @@
 from transformers.utils.versions import require_version
 
 from optimum.intel.neural_compressor import INCModelForCausalLM, INCQuantizer, INCTrainer
-from optimum.intel.utils.import_utils import ITREX_IMPORT_ERROR, is_itrex_available
 
 
-if is_itrex_available():
-    from intel_extension_for_transformers.transformers.utils.config import GPTQConfig, RtnConfig
-
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
@@ -203,12 +200,8 @@ class OptimizationArguments:
         metadata={"help": "Whether or not to verify the loading of the quantized model."},
     )
     bits: str = field(
-        default="4",
-        metadata={"help": "Bits number of weight for weight only quantization. 1~8 bits."},
-    )
-    weight_dtype: str = field(
-        default="int4_clip",
-        metadata={"help": "weight dtype for weight only quantization."},
+        default=4,
+        metadata={"help": "Bits number of weight for weight only quantization. only support 4 bits now."},
     )
     group_size: int = field(
         default=-1,
@@ -223,7 +216,6 @@ class OptimizationArguments:
         metadata={"help": "Scheme for weight only quantization. Choose from 'sym' and 'asym'."},
     )
     quantization_methodology: str = field(
-        choices=["rtn", "gptq"],
         default="rtn",
         metadata={"help": "Quantization methodology for weight only quantization. Choose from 'rtn' and 'gptq'."},
     )
@@ -249,6 +241,11 @@ class OptimizationArguments:
         metadata={"help": "Calibration dataset sequence max length, this should align with your model config"},
     )
 
+    def __post_init__(self):
+        woq_algorithms = ["rtn", "gptq"]
+        if self.quantization_methodology not in woq_algorithms:
+            raise ValueError(f"Value must be one of {woq_algorithms}, got {self.quantization_methodology}")
+
 
 @dataclass
 class DataTrainingArguments:
@@ -655,13 +652,11 @@ def compute_metrics(eval_preds):
             else:
                 recipes = {}
             if optim_args.quantization_approach == "weight_only":
-                if not is_itrex_available():
-                    raise ImportError(ITREX_IMPORT_ERROR.format("WeightOnly quantization"))
                 if optim_args.apply_pruning or optim_args.apply_distillation:
                     raise ValueError("Weight only quantization and pruning or distillation cannot be combined.")
 
                 algorithm_args = {
-                    "weight_dtype": optim_args.weight_dtype,
+                    "bits": optim_args.bits,
                     "sym": optim_args.weight_only_scheme == "sym",
                     "group_size": optim_args.group_size,
                 }
@@ -756,10 +751,10 @@ def compute_metrics(eval_preds):
         trainer.save_metrics("train", metrics)
         trainer.save_state()
 
-    if optim_args.apply_quantization and optim_args.quantization_approach in {"static", "dynamic", "weight_only"}:
+    if optim_args.apply_quantization and optim_args.quantization_approach in {"static", "dynamic"}:
         model = trainer.model if isinstance(trainer.model, PreTrainedModel) else trainer.model._model
         quantizer = INCQuantizer.from_pretrained(model)
-        if optim_args.quantization_approach in ["static", "weight_only"]:
+        if optim_args.quantization_approach == "static":
             num_calibration_samples = min(len(train_dataset), optim_args.num_calibration_samples)
             train_dataset = train_dataset.select(range(num_calibration_samples))
             quantization_config.calibration_sampling_size = num_calibration_samples
@@ -776,6 +771,19 @@ def compute_metrics(eval_preds):
         )
         trainer.model = quantizer._quantized_model
 
+    if optim_args.apply_quantization and optim_args.quantization_approach == "weight_only":
+        model = trainer.model if isinstance(trainer.model, PreTrainedModel) else trainer.model._model
+        num_calibration_samples = min(len(train_dataset), optim_args.num_calibration_samples)
+        train_dataset = train_dataset.select(range(num_calibration_samples))
+        quantization_config.calibration_sampling_size = num_calibration_samples
+        quantized_model = INCModelForCausalLM.from_pretrained(
+            model_args.model_name_or_path, quantization_config=quantization_config
+        )
+        if hasattr(quantization_config, "tokenizer"):
+            quantization_config.tokenizer.save_pretrained(training_args.output_dir)
+            quantized_model.save_pretrained(training_args.output_dir)
+        trainer.model = quantized_model
+
     if optim_args.apply_quantization and optim_args.verify_loading:
         loaded_model = INCModelForCausalLM.from_pretrained(training_args.output_dir)
         tokens = tokenizer("This is a sample input", return_tensors="pt")

diff --git a/optimum/intel/neural_compressor/modeling_base.py b/optimum/intel/neural_compressor/modeling_base.py
@@ -23,6 +23,8 @@
 from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
 from huggingface_hub.utils import EntryNotFoundError
+from neural_compressor.transformers import GPTQConfig, RtnConfig
+from neural_compressor.transformers.models.modeling_auto import _BaseINCAutoModelClass
 from neural_compressor.utils.pytorch import load
 from transformers import (
     AutoConfig,
@@ -47,8 +49,9 @@
 from optimum.intel.generation import BaseModelForCausalLM
 
 from ...modeling_base import OptimizedModel
-from ..utils.import_utils import _torch_version, is_itrex_available, is_torch_version
+from ..utils.import_utils import _torch_version, is_torch_version
 from .configuration import INCConfig
+from .quantization import _weight_only_quantization
 from .utils import QUANTIZATION_CONFIG_NAME
 
 
@@ -122,8 +125,85 @@ def _from_pretrained(
                 raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
             token = use_auth_token
 
+        quantization_config = kwargs.pop("quantization_config", None)
         model_path = Path(model_id)
         is_local = model_path.is_dir()
+
+        # ITREX compatibility
+        quantization_config_path = None
+        if is_local:
+            quantization_config_path = model_path / subfolder / QUANTIZATION_CONFIG_NAME
+        else:
+            try:
+                quantization_config_path = hf_hub_download(
+                    repo_id=model_id,
+                    filename=QUANTIZATION_CONFIG_NAME,
+                    subfolder=subfolder,
+                    token=token,
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    local_files_only=local_files_only,
+                )
+            except EntryNotFoundError:
+                pass
+        if quantization_config_path and Path(quantization_config_path).is_file():
+            algorithm = getattr(quantization_config, "quant_method", None)
+            if algorithm in {"rtn", "gptq", "awq", "autoround"}:
+                raise ValueError(
+                    "This model was obtained through ITREX quantization, support for ITREX models is deprecated since neural-compressor v3.0. "
+                    "To load this model please downgrade both optimum-intel and neural-compressor."
+                )
+                # quantization_config = PretrainedConfig.from_pretrained(quantization_config_path)
+                # config.quantization_config = quantization_config.to_dict()
+
+        if hasattr(config, "quantization_config"):
+            if config.quantization_config is None:
+                raise ValueError(
+                    "The loading of `quantization_config` failed, to load this model please make sure the config is compatible"
+                )
+            else:
+                try:
+                    logger.info(
+                        "The weight only quantized model loading only supports the same format as GPTQ, such as https://huggingface.co/TheBloke/Llama-2-7B-Chat-GPTQ/tree/main."
+                    )
+                    _BaseINCAutoModelClass.ORIG_MODEL = cls.auto_model_class
+                    model = _BaseINCAutoModelClass.load_low_bit(
+                        model_id,
+                        subfolder=subfolder,
+                        revision=revision,
+                        cache_dir=cache_dir,
+                        token=token,
+                        local_files_only=local_files_only,
+                        force_download=force_download,
+                        trust_remote_code=trust_remote_code,
+                        config=config,
+                        **kwargs,
+                    )
+                    logger.info("Saved low bit model loading successfully. Other input args " "will be ignored.")
+                    return model
+                except Exception as e:
+                    raise RuntimeError(f"The quantized model cannot be loaded. Detailed error: {e}")
+        if isinstance(quantization_config, (RtnConfig, GPTQConfig)):
+            logger.info(
+                "The quantized model parameters will be saved in the same format as GPTQ, here is the sample model https://huggingface.co/TheBloke/Llama-2-7B-Chat-GPTQ/tree/main for details."
+            )
+            model = _weight_only_quantization(
+                cls.auto_model_class,
+                model_id,
+                quantization_config=quantization_config,
+                token=token,
+                revision=revision,
+                force_download=force_download,
+                cache_dir=cache_dir,
+                local_files_only=local_files_only,
+                subfolder=subfolder,
+                trust_remote_code=trust_remote_code,
+                **kwargs,
+            )
+
+            return cls(model, config=config, model_save_dir=None, **kwargs).model
+
         model_cache_path = None
         inc_config = None
         msg = None
@@ -165,52 +245,6 @@ def _from_pretrained(
 
         model_save_dir = Path(model_cache_path).parent
 
-        if is_itrex_available():
-            quantization_config_path = None
-            if is_local:
-                quantization_config_path = model_path / subfolder / QUANTIZATION_CONFIG_NAME
-            else:
-                try:
-                    quantization_config_path = hf_hub_download(
-                        repo_id=model_id,
-                        filename=QUANTIZATION_CONFIG_NAME,
-                        subfolder=subfolder,
-                        token=token,
-                        revision=revision,
-                        cache_dir=cache_dir,
-                        force_download=force_download,
-                        local_files_only=local_files_only,
-                    )
-                except EntryNotFoundError:
-                    pass
-
-            if quantization_config_path and Path(quantization_config_path).is_file():
-                quantization_config = PretrainedConfig.from_pretrained(quantization_config_path)
-                algorithm = getattr(quantization_config, "quant_method", None)
-                if algorithm in {"rtn", "gptq", "awq", "autoround"}:
-                    from intel_extension_for_transformers.transformers.modeling.modeling_auto import (
-                        _BaseQBitsAutoModelClass,
-                    )
-
-                    _BaseQBitsAutoModelClass.ORIG_MODEL = cls.auto_model_class
-
-                    model = _BaseQBitsAutoModelClass.from_pretrained(
-                        pretrained_model_name_or_path=model_id,
-                        token=token,
-                        revision=revision,
-                        force_download=force_download,
-                        cache_dir=cache_dir,
-                        local_files_only=local_files_only,
-                        subfolder=subfolder,
-                        trust_remote_code=trust_remote_code,
-                        use_neural_speed=False,
-                        **kwargs,
-                    )
-
-                    return cls(
-                        model, config=config, model_save_dir=model_save_dir, q_config=quantization_config, **kwargs
-                    )
-
         try:
             inc_config = INCConfig.from_pretrained(model_id, subfolder=subfolder, revision=revision)
             if not is_torch_version("==", inc_config.torch_version):
@@ -254,7 +288,7 @@ def _from_pretrained(
 
     def _save_pretrained(self, save_directory: Union[str, Path]):
         if isinstance(self.model, torch.nn.Module):
-            # For ITREX model
+            # For INC weight only model
             if isinstance(self._q_config, PretrainedConfig):
                 self._q_config.to_json_file(os.path.join(save_directory, QUANTIZATION_CONFIG_NAME))
                 self.model.save_pretrained(save_directory)