Merge branch 'main' into del-convert-tokenizer-flag

Author: echarlaix

.github/workflows/test_inc.yml

@@ -32,11 +32,17 @@ jobs:
         python -m pip install --upgrade pip
         pip install cmake
         pip install py-cpuinfo
-        pip install torch==2.1.0 torchaudio==2.1.0 torchvision==0.16 --extra-index-url https://download.pytorch.org/whl/cpu
         pip install .[neural-compressor,diffusers,tests]
-        pip install intel-extension-for-pytorch==2.1.100
-        pip install intel-extension-for-transformers==1.3.2
+        pip install intel-extension-for-transformers
         pip install peft
+
     - name: Test with Pytest
       run: |
-        pytest tests/neural_compressor/
+        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 torch==2.1.0 torchaudio==2.1.0 torchvision==0.16 --extra-index-url https://download.pytorch.org/whl/cpu
+        pip install intel-extension-for-pytorch==2.1.100
+        pytest tests/neural_compressor/test_ipex.py
+

README.md

@@ -78,12 +78,18 @@ It is possible to export your model to the [OpenVINO IR](https://docs.openvino.a
 optimum-cli export openvino --model gpt2 ov_model
 ```
 
-You can also apply 8-bit weight-only quantization when exporting your model : the model linear and embedding weights will be quantized to INT8, the activations will be kept in floating point precision.
+You can also apply 8-bit weight-only quantization when exporting your model : the model linear, embedding and convolution weights will be quantized to INT8, the activations will be kept in floating point precision.
 
 ```plain
 optimum-cli export openvino --model gpt2 --weight-format int8 ov_model
 ```
 
+Quantization in hybrid mode can be applied to Stable Diffusion pipeline during model export. This involves applying hybrid post-training quantization to the UNet model and weight-only quantization for the rest of the pipeline components. In the hybrid mode, weights in MatMul and Embedding layers are quantized, as well as activations of other layers.
+
+```plain
+optimum-cli export openvino --model stabilityai/stable-diffusion-2-1 --dataset conceptual_captions --weight-format int8 ov_model
+```
+
 To apply quantization on both weights and activations, you can find more information in the [documentation](https://huggingface.co/docs/optimum/main/en/intel/optimization_ov).
 
 #### Inference:

examples/neural_compressor/language-modeling/run_clm.py

@@ -64,8 +64,7 @@
 
 
 if is_intel_extension_for_transformers_available():
-    from intel_extension_for_transformers.transformers.utils.config import WeightOnlyQuantConfig
-
+    from intel_extension_for_transformers.transformers.utils.config import GPTQConfig, RtnConfig
 
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
@@ -227,8 +226,9 @@ class OptimizationArguments:
         metadata={"help": "Scheme for weight only quantization. Choose from 'sym' and 'asym'."},
     )
     quantization_methodology: str = field(
-        default="RTN",
-        metadata={"help": "Quantization methodology for weight only quantization. Choose from 'RTN' and 'GPTQ'."},
+        choices=["rtn", "gptq"],
+        default="rtn",
+        metadata={"help": "Quantization methodology for weight only quantization. Choose from 'rtn' and 'gptq'."},
     )
     damp_percent: float = field(
         default=0.01,
@@ -662,22 +662,23 @@ def compute_metrics(eval_preds):
                     raise ImportError(INTEL_EXTENSION_FOR_TRANSFORMERS_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.")
-                if optim_args.quantization_methodology == "GPTQ":
-                    algorithm_args = {
-                        "act_order": False,
-                        "percdamp": optim_args.damp_percent,
-                        "block_size": optim_args.gptq_block_size,
-                        "nsamples": optim_args.num_calibration_samples,
-                        "use_max_length": optim_args.use_max_length,
-                        "pad_max_length": optim_args.pad_max_length,
-                    }
-                quantization_config = WeightOnlyQuantConfig(
-                    weight_dtype=optim_args.weight_dtype,
-                    group_size=optim_args.group_size,
-                    scheme=optim_args.weight_only_scheme,
-                    algorithm=optim_args.quantization_methodology,
-                    algorithm_args=algorithm_args if optim_args.quantization_methodology == "GPTQ" else None,
-                )
+
+                algorithm_args = {
+                    "weight_dtype": optim_args.weight_dtype,
+                    "sym": optim_args.weight_only_scheme == "sym",
+                    "group_size": optim_args.group_size,
+                }
+
+                if optim_args.quantization_methodology == "gptq":
+                    quantization_config = GPTQConfig(
+                        damp_percent=optim_args.damp_percent,
+                        nsamples=optim_args.num_calibration_samples,
+                        blocksize=optim_args.gptq_block_size,
+                        **algorithm_args,
+                    )
+                else:
+                    quantization_config = RtnConfig(**algorithm_args)
+
             else:
                 quantization_config = PostTrainingQuantConfig(
                     approach=optim_args.quantization_approach, recipes=recipes

optimum/commands/export/openvino.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING, Optional
 
 from ...exporters import TasksManager
+from ...intel.utils.import_utils import DIFFUSERS_IMPORT_ERROR, is_diffusers_available
 from ..base import BaseOptimumCLICommand, CommandInfo
 
 
@@ -104,6 +105,16 @@ def parse_args_openvino(parser: "ArgumentParser"):
         default=None,
         help=("The group size to use for quantization. Recommended value is 128 and -1 uses per-column quantization."),
     )
+    optional_group.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        help=(
+            "The dataset used for data-aware compression or quantization with NNCF. "
+            "You can use the one from the list ['wikitext2','c4','c4-new','ptb','ptb-new'] for LLLMs "
+            "or ['conceptual_captions','laion/220k-GPT4Vision-captions-from-LIVIS','laion/filtered-wit'] for diffusion models."
+        ),
+    )
     optional_group.add_argument(
         "--disable-stateful",
         action="store_true",
@@ -200,23 +211,64 @@ def run(self):
                 )
                 quantization_config["sym"] = "asym" not in self.args.weight_format
                 quantization_config["group_size"] = 128 if "128" in self.args.weight_format else 64
+            quantization_config["dataset"] = self.args.dataset
             ov_config = OVConfig(quantization_config=quantization_config)
 
-        if self.args.convert_tokenizer:
-            logger.warning("`--convert-tokenizer` option is deprecated. Tokenizer will be converted by default.")
-
-        # TODO : add input shapes
-        main_export(
-            model_name_or_path=self.args.model,
-            output=self.args.output,
-            task=self.args.task,
-            framework=self.args.framework,
-            cache_dir=self.args.cache_dir,
-            trust_remote_code=self.args.trust_remote_code,
-            pad_token_id=self.args.pad_token_id,
-            ov_config=ov_config,
-            stateful=not self.args.disable_stateful,
-            convert_tokenizer=not self.args.disable_convert_tokenizer,
-            library_name=self.args.library
-            # **input_shapes,
-        )
+        library_name = TasksManager.infer_library_from_model(self.args.model)
+
+        if (
+            library_name == "diffusers"
+            and ov_config
+            and ov_config.quantization_config
+            and ov_config.quantization_config.dataset is not None
+        ):
+            if not is_diffusers_available():
+                raise ValueError(DIFFUSERS_IMPORT_ERROR.format("Export of diffusers models"))
+
+            from diffusers import DiffusionPipeline
+
+            diffusers_config = DiffusionPipeline.load_config(self.args.model)
+            class_name = diffusers_config.get("_class_name", None)
+
+            if class_name == "LatentConsistencyModelPipeline":
+                from optimum.intel import OVLatentConsistencyModelPipeline
+
+                model_cls = OVLatentConsistencyModelPipeline
+
+            elif class_name == "StableDiffusionXLPipeline":
+                from optimum.intel import OVStableDiffusionXLPipeline
+
+                model_cls = OVStableDiffusionXLPipeline
+            elif class_name == "StableDiffusionPipeline":
+                from optimum.intel import OVStableDiffusionPipeline
+
+                model_cls = OVStableDiffusionPipeline
+            else:
+                raise NotImplementedError(f"Quantization in hybrid mode isn't supported for class {class_name}.")
+
+            model = model_cls.from_pretrained(
+                self.args.model, export=True, quantization_config=ov_config.quantization_config
+            )
+            model.save_pretrained(self.args.output)
+
+        else:
+            if self.args.convert_tokenizer:
+                logger.warning("`--convert-tokenizer` option is deprecated. Tokenizer will be converted by default.")
+
+            # TODO : add input shapes
+            main_export(
+                model_name_or_path=self.args.model,
+                output=self.args.output,
+                task=self.args.task,
+                framework=self.args.framework,
+                cache_dir=self.args.cache_dir,
+                trust_remote_code=self.args.trust_remote_code,
+                pad_token_id=self.args.pad_token_id,
+                ov_config=ov_config,
+                stateful=not self.args.disable_stateful,
+                convert_tokenizer=not self.args.disable_convert_tokenizer,
+                library_name=library_name,
+                # **input_shapes,
+            )
+            
+

optimum/exporters/openvino/__main__.py

@@ -76,7 +76,7 @@ def main_export(
         model_name_or_path (`str`):
             Model ID on huggingface.co or path on disk to the model repository to export.
         output (`Union[str, Path]`):
-            Path indicating the directory where to store the generated ONNX model.
+            Path indicating the directory where to store the generated OpenVINO model.
 
         > Optional parameters
 

optimum/intel/neural_compressor/modeling_base.py

@@ -67,11 +67,6 @@
 """
 
 
-if is_intel_extension_for_transformers_available():
-    from intel_extension_for_transformers.transformers.modeling import AutoModelForCausalLM as ITREX_WOQ_MODEL
-    from intel_extension_for_transformers.transformers.utils import WeightOnlyQuantConfig
-
-
 class INCModel(OptimizedModel):
     auto_model_class = AutoModel
     export_feature = "feature-extraction"
@@ -142,15 +137,16 @@ def _from_pretrained(
         msg = None
         if is_intel_extension_for_transformers_available():
             try:
-                quantization_config = WeightOnlyQuantConfig.from_pretrained(model_id)
-                algorithm = getattr(quantization_config, "algorithm", None)
-                if algorithm is not None and quantization_config.algorithm.lower() in {
-                    "rtn",
-                    "gptq",
-                    "awq",
-                    "autoaround",
-                }:
-                    return ITREX_WOQ_MODEL.from_pretrained(
+                quantization_config = PretrainedConfig.from_pretrained(model_save_dir / "quantize_config.json")
+                algorithm = getattr(quantization_config, "quant_method", None)
+                if algorithm in {"rtn", "gptq", "awq", "autoaround"}:
+                    from intel_extension_for_transformers.transformers.modeling.modeling_auto import (
+                        _BaseQBitsAutoModelClass,
+                    )
+
+                    _BaseQBitsAutoModelClass.ORIG_MODEL = cls.auto_model_class
+
+                    return _BaseQBitsAutoModelClass.from_pretrained(
                         pretrained_model_name_or_path=model_id,
                         use_auth_token=use_auth_token,
                         revision=revision,