Merge branch 'main' into task-to-padding-side

Author: apaniukov

.github/workflows/dockerfile_sanity.yml

@@ -0,0 +1,44 @@
+name: Build and Test Docker Image
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - 'docker/Dockerfile.intel'
+ 
+  pull_request:
+    branches:
+      - main
+    paths:
+      - 'docker/Dockerfile.intel'
+
+jobs:
+  build_and_run:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+
+    - name: Set up Docker Buildx
+      uses: docker/setup-buildx-action@v3
+
+    - name: Build and Run Docker Image
+      run: |
+        IMAGE_NAME="intel_image:latest"
+        docker build -f docker/Dockerfile.intel -t $IMAGE_NAME .
+        if [ $? -ne 0 ]; then
+          echo "Docker image build failed."
+          exit 1
+        fi
+        CONTAINER_ID=$(docker run -d $IMAGE_NAME tail -f /dev/null)
+        if docker inspect -f '{{.State.Running}}' $CONTAINER_ID 2>/dev/null | grep -q 'true'; then
+          echo "Container is running."
+        else
+          echo "Container failed to start."
+          docker logs $CONTAINER_ID 2>/dev/null || echo "No container ID found."
+          exit 1
+        fi
+        docker stop $CONTAINER_ID
+        docker rm $CONTAINER_ID

.github/workflows/test_openvino_notebooks.yml

@@ -40,8 +40,8 @@ jobs:
         # ffmpeg, torchaudio and pillow are required for image classification and audio classification pipelines
         sudo apt-get install ffmpeg
         pip install torch torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
-        pip install ".[tests, openvino]" nbval
         pip install -r notebooks/openvino/requirements.txt
+        pip install .[tests,openvino] nbval
 
     - run: free -h
     - run: lscpu

README.md

@@ -72,7 +72,7 @@ Below are examples of how to use OpenVINO and its [NNCF](https://docs.openvino.a
 
 #### Export:
 
-It is possible to export your model to the [OpenVINO IR](https://docs.openvino.ai/2024/documentation/openvino-ir-format.html) format with the CLI :
+It is also possible to export your model to the [OpenVINO IR](https://docs.openvino.ai/2024/documentation/openvino-ir-format.html) format with the CLI :
 
 ```plain
 optimum-cli export openvino --model gpt2 ov_model

docker/Dockerfile.intel

@@ -37,6 +37,7 @@ ARG TORCHVISION_VERSION=0.18.1+cpu
 ARG TORCHAUDIO_VERSION=2.3.1+cpu
 
 RUN python3 -m pip install --no-cache-dir \
+    intel-openmp \
     torch==${PYTORCH_VERSION}+cpu \
     torchvision==${TORCHVISION_VERSION} \
     torchaudio==${TORCHAUDIO_VERSION} \

docs/source/openvino/export.mdx

@@ -9,6 +9,8 @@ specific language governing permissions and limitations under the License.
 
 # Export your model
 
+To export a [model](https://huggingface.co/docs/optimum/main/en/intel/openvino/models) hosted on the [Hub](https://huggingface.co/models) you can use our [space](https://huggingface.co/spaces/echarlaix/openvino-export). After conversion, a repository will be pushed under your namespace, this repository can be either public or private.
+
 ## Using the CLI
 
 To export your model to the [OpenVINO IR](https://docs.openvino.ai/2024/documentation/openvino-ir-format.html) format with the CLI :

docs/source/openvino/optimization.mdx

@@ -64,7 +64,7 @@ model = OVModelForCausalLM.from_pretrained(model_id, quantization_config=quantiz
 You can tune quantization parameters to achieve a better performance accuracy trade-off as follows:
 
 ```python
-quantization_config = OVWeightQuantizationConfig(bits=4, sym=False, ratio=0.8, dataset="ptb")
+quantization_config = OVWeightQuantizationConfig(bits=4, sym=False, ratio=0.8, dataset="wikitext2")
 ```
 
 By default the quantization scheme will be [asymmetric](https://github.com/openvinotoolkit/nncf/blob/develop/docs/usage/training_time_compression/other_algorithms/LegacyQuantization.md#asymmetric-quantization), to make it [symmetric](https://github.com/openvinotoolkit/nncf/blob/develop/docs/usage/training_time_compression/other_algorithms/LegacyQuantization.md#symmetric-quantization) you can add `sym=True`.

examples/openvino/audio-classification/requirements.txt

@@ -1,5 +1,5 @@
-datasets>=1.14.0
+datasets>=1.14.0,<2.20.0
 evaluate
 librosa
 torchaudio
-accelerate
+accelerate

examples/openvino/image-classification/requirements.txt

@@ -1,4 +1,4 @@
-datasets >= 1.8.0
+datasets>=1.14.0,<2.20.0
 torch >= 1.9.0
 torchvision>=0.6.0
 evaluate

examples/openvino/question-answering/requirements.txt

@@ -1,4 +1,4 @@
-datasets >= 1.8.0
+datasets>=1.14.0,<2.20.0
 torch >= 1.9.0
 evaluate
 accelerate

examples/openvino/text-classification/requirements.txt

@@ -1,4 +1,4 @@
-datasets >= 1.8.0
+datasets>=1.14.0,<2.20.0
 sentencepiece != 0.1.92
 scipy
 scikit-learn

optimum/commands/export/openvino.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 """Defines the command line for the export with OpenVINO."""
 
-import json
 import logging
 import sys
 from pathlib import Path
@@ -213,19 +212,7 @@ 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_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:
-                return _DEFAULT_4BIT_CONFIGS[model_id_or_path]
-            if "transformers" in library_name and (Path(model_id_or_path) / "config.json").exists():
-                with (Path(model_id_or_path) / "config.json").open("r") as config_f:
-                    config = json.load(config_f)
-                    original_model_name = config.get("_name_or_path", "")
-                if original_model_name in _DEFAULT_4BIT_CONFIGS:
-                    return _DEFAULT_4BIT_CONFIGS[original_model_name]
-
-            return _DEFAULT_4BIT_CONFIG
+        from ...intel.openvino.configuration import _DEFAULT_4BIT_CONFIG, OVConfig, get_default_int4_config
 
         if self.args.library is None:
             # TODO: add revision, subfolder and token to args
@@ -260,7 +247,7 @@ def _get_default_int4_config(model_id_or_path, library_name):
                 and self.args.awq is None
                 and self.args.sensitivity_metric is None
             ):
-                quantization_config = _get_default_int4_config(self.args.model, library_name)
+                quantization_config = get_default_int4_config(self.args.model)
             else:
                 quantization_config = {
                     "bits": 8 if is_int8 else 4,

optimum/exporters/openvino/model_configs.py

@@ -24,6 +24,7 @@
     FalconOnnxConfig,
     GemmaOnnxConfig,
     LlamaOnnxConfig,
+    MistralOnnxConfig,
     MPTOnnxConfig,
     PhiOnnxConfig,
     UNetOnnxConfig,
@@ -53,11 +54,13 @@
     InternLMModelPatcher,
     JaisModelPatcher,
     LlamaModelPatcher,
+    MistralModelPatcher,
     MixtralModelPatcher,
     MPTModelPatcher,
     PersimmonModelPatcher,
     Phi3ModelPatcher,
     QwenModelPatcher,
+    UpdateCausalMaskModelPatcher,
     XverseModelPatcher,
 )
 
@@ -117,6 +120,11 @@ class Qwen2OpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
     DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
 
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return UpdateCausalMaskModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 @register_in_tasks_manager("qwen2-moe", *["text-generation", "text-generation-with-past"], library_name="transformers")
 class Qwen2MoEOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
@@ -126,6 +134,11 @@ class Qwen2MoEOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
     DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
 
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return UpdateCausalMaskModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 @register_in_tasks_manager("minicpm", *["text-generation", "text-generation-with-past"], library_name="transformers")
 class MiniCPMOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
@@ -144,6 +157,11 @@ class StableLMOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
     DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
 
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return UpdateCausalMaskModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 class ChatGLM2DummyPastKeyValuesGenerator(DummyPastKeyValuesGenerator):
     def __init__(
@@ -466,6 +484,11 @@ class Starcoder2OpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
     DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
 
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return UpdateCausalMaskModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 @register_in_tasks_manager("internlm2", *["text-generation", "text-generation-with-past"], library_name="transformers")
 class InternLM2OpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
@@ -530,6 +553,24 @@ def patch_model_for_export(
         return Phi3ModelPatcher(self, model, model_kwargs=model_kwargs)
 
 
+@register_in_tasks_manager(
+    "phi",
+    *[
+        "feature-extraction",
+        "feature-extraction-with-past",
+        "text-generation",
+        "text-generation-with-past",
+        "text-classification",
+    ],
+    library_name="transformers",
+)
+class PhiOpenVINOConfig(PhiOnnxConfig):
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return UpdateCausalMaskModelPatcher(self, model, model_kwargs=model_kwargs)
+
+
 class OVFalconDummyPastKeyValuesGenerator(FalconDummyPastKeyValuesGenerator):
     def __init__(
         self,
@@ -839,3 +880,21 @@ def patch_model_for_export(
             )
 
         return ArcticModelPatcher(self, model, model_kwargs=model_kwargs)
+
+
+@register_in_tasks_manager(
+    "mistral",
+    *[
+        "feature-extraction",
+        "feature-extraction-with-past",
+        "text-generation",
+        "text-generation-with-past",
+        "text-classification",
+    ],
+    library_name="transformers",
+)
+class MistralOpenVINOConfig(MistralOnnxConfig):
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return MistralModelPatcher(self, model, model_kwargs=model_kwargs)