Merge branch 'openvinotoolkit:develop' into add-tensorflow-tensor-support

Author: darshil929

.github/workflows/api_changes_check.yml

@@ -26,7 +26,7 @@ jobs:
       issues: write
     steps:
       - name: Download built HTML doc as artifact from previous step
-        uses: actions/download-artifact@cc203385981b70ca67e1cc392babf9cc229d5806 # v4.1.9
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: html_doc_artifact
       - run: |
@@ -63,7 +63,7 @@ jobs:
           echo '{"pr_number": "${{ github.event.pull_request.number }}", "action": "none"}' > api_status.json
 
       - name: Upload artifact
-        uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 #v4.6.1
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 #v4.6.2
         with:
           name: api_status
           path: api_status.json

.github/workflows/build_and_publish_doc.yml

@@ -28,7 +28,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: Download HTML doc build artifact
-        uses: actions/download-artifact@cc203385981b70ca67e1cc392babf9cc229d5806 # v4.1.9
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: html_doc_artifact
       - name: Extract artifact
@@ -38,7 +38,7 @@ jobs:
           rm artifact.tar
 
       - name: Download schema doc build artifact
-        uses: actions/download-artifact@cc203385981b70ca67e1cc392babf9cc229d5806 # v4.1.9
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: schema_doc_artifact
           path: html_build/html

.github/workflows/build_html_doc.yml

@@ -29,7 +29,7 @@ jobs:
       - name: Archive built HTMLs
         shell: bash
         run: tar -czf artifact.tar html_build/html
-      - uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 #v4.6.1
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 #v4.6.2
         with:
           name: html_doc_artifact
           path: artifact.tar

.github/workflows/build_schema_page.yml

@@ -31,7 +31,7 @@ jobs:
         run: tar -czf artifact.tar schema
 
       - name: Upload result as artifact
-        uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 #v4.6.1
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 #v4.6.2
         with:
           name: schema_doc_artifact
           path: artifact.tar

.github/workflows/conformance_weight_compression.yml

@@ -62,7 +62,7 @@ jobs:
         run: column -s, -t < tmp/results.csv || echo "no file"
       - name: Upload artifact
         if: ${{ !cancelled() }}
-        uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 #v4.6.1
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 #v4.6.2
         with:
           name: wc_results_${{ matrix.group }}
           path: tmp/results.csv

.github/workflows/nightly.yml

@@ -21,7 +21,7 @@ jobs:
     timeout-minutes: 10
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - uses: AlexanderDokuchaev/md-dead-link-check@c7210ef8a38c194a119834e39d212387d19b512c # v1.1.0
+      - uses: AlexanderDokuchaev/md-dead-link-check@d5a37e0b14e5918605d22b34562532762ccb2e47 # v1.2.0
 
   tensorflow:
     runs-on: ubuntu-latest-8-cores

.github/workflows/pre-commit-linters.yml

@@ -24,5 +24,5 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - uses: AlexanderDokuchaev/md-dead-link-check@c7210ef8a38c194a119834e39d212387d19b512c # v1.1.0
+      - uses: AlexanderDokuchaev/md-dead-link-check@d5a37e0b14e5918605d22b34562532762ccb2e47 # v1.2.0
 

.github/workflows/sdl.yml

@@ -42,11 +42,11 @@ jobs:
         with:
           lfs: true
       - name: Initialize CodeQL
-        uses: github/codeql-action/init@b56ba49b26e50535fa1e7f7db0f4f7b4bf65d80d # v3.28.10
+        uses: github/codeql-action/init@5f8171a638ada777af81d42b55959a643bb29017 # v3.28.12
         with:
           languages: python
       - name: Perform CodeQL Analysis
-        uses: github/codeql-action/analyze@b56ba49b26e50535fa1e7f7db0f4f7b4bf65d80d # v3.28.10
+        uses: github/codeql-action/analyze@5f8171a638ada777af81d42b55959a643bb29017 # v3.28.12
         with:
           category: "/language:python"
 
@@ -66,7 +66,7 @@ jobs:
           mv "report.pdf" "codeql_nncf_report_${DATE}_${REF_NAME//\//-}_${{ github.sha }}.pdf"
       - name: Upload CodeQL Artifacts
         if: ${{ github.event_name != 'pull_request' }}
-        uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 #v4.6.1
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 #v4.6.2
         with:
           name: codeql-scan-results
           path: "./codeql*.pdf"
@@ -81,7 +81,7 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
       - name: Run trivy
-        uses: aquasecurity/trivy-action@18f2510ee396bbf400402947b394f2dd8c87dbb0 # v0.29.0
+        uses: aquasecurity/trivy-action@6c175e9c4083a92bbca2f9724c8a5e33bc2d97a5 # v0.30.0
         with:
           scan-type: "fs"
           scan-ref: .
@@ -102,7 +102,7 @@ jobs:
           mv "trivy_report.html" "trivy_report_${DATE}_${REF_NAME//\//-}_${{ github.sha }}.html"
       - name: Upload Scan Results
         if: ${{ !cancelled() }}
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 #v4.6.0
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 #v4.6.2
         with:
           name: trivy-report
           path: "./trivy_report*.html"

README.md

@@ -245,6 +245,7 @@ Here is an example of Accuracy Aware Quantization pipeline where model weights a
 
 ```python
 import nncf
+import nncf.torch
 import torch
 from torchvision import datasets, models
 
@@ -271,7 +272,7 @@ quantized_model = nncf.quantize(model, calibration_dataset)
 # Save quantization modules and the quantized model parameters
 checkpoint = {
     'state_dict': model.state_dict(),
-    'nncf_config': model.nncf.get_config(),
+    'nncf_config': nncf.torch.get_config(model),
     ... # the rest of the user-defined objects to save
 }
 torch.save(checkpoint, path_to_checkpoint)

docs/usage/training_time_compression/quantization_aware_training/Usage.md

@@ -75,17 +75,19 @@ ov_quantized_model = ov.convert_model(stripped_model)
 
 The complete information about compression is defined by a compressed model and a NNCF config.
 The model characterizes the weights and topology of the network. The NNCF config - how to restore additional modules introduced by NNCF.
-The NNCF config can be obtained by `quantized_model.nncf.get_config()` on saving and passed to the
+The NNCF config can be obtained by `nncf.torch.get_config` on saving and passed to the
 `nncf.torch.load_from_config` helper function to load additional modules from the given NNCF config.
 The quantized model saving allows to load quantized modules to the target model in a new python process and
 requires only example input for the target module, corresponding NNCF config and the quantized model state dict.
 
 ```python
+import nncf.torch
+
 # save part
 quantized_model = nncf.quantize(model, calibration_dataset)
 checkpoint = {
-    'state_dict':quantized_model.state_dict(),
-    'nncf_config': quantized_model.nncf.get_config(),
+    'state_dict': quantized_model.state_dict(),
+    'nncf_config': nncf.torch.get_config(quantized_model),
     ...
 }
 torch.save(checkpoint, path)

examples/quantization_aware_training/torch/resnet18/main.py

@@ -278,11 +278,11 @@ def transform_fn(data_item):
         print(f"Train epoch: {epoch}")
         train_epoch(train_loader, quantized_model, criterion, optimizer, device=device)
         acc1_int8 = validate(val_loader, quantized_model, device)
-        print(f"Accyracy@1 of INT8 model after {epoch} epoch finetuning: {acc1_int8:.3f}")
+        print(f"Accuracy@1 of INT8 model after {epoch} epoch finetuning: {acc1_int8:.3f}")
         # Save the compression checkpoint for model with the best accuracy metric.
         if acc1_int8 > acc1_int8_best:
             state_dict = quantized_model.state_dict()
-            compression_config = quantized_model.nncf.get_config()
+            compression_config = nncf.torch.get_config(quantized_model)
             torch.save(
                 {
                     "model_state_dict": state_dict,

nncf/__init__.py

@@ -40,6 +40,7 @@
 from nncf.parameters import ModelType as ModelType
 from nncf.parameters import QuantizationMode as QuantizationMode
 from nncf.parameters import SensitivityMetric as SensitivityMetric
+from nncf.parameters import StripFormat as StripFormat
 from nncf.parameters import TargetDevice as TargetDevice
 from nncf.quantization import QuantizationPreset as QuantizationPreset
 from nncf.quantization import compress_weights as compress_weights

nncf/api/compression.py

@@ -19,6 +19,7 @@
 from nncf.common.statistics import NNCFStatistics
 from nncf.common.utils.api_marker import api
 from nncf.common.utils.backend import copy_model
+from nncf.parameters import StripFormat
 
 TModel = TypeVar("TModel")
 
@@ -236,14 +237,17 @@ def statistics(self, quickly_collected_only: bool = False) -> NNCFStatistics:
             need to keep track of statistics on each training batch/step/iteration.
         """
 
-    def strip_model(self, model: TModel, do_copy: bool = False) -> TModel:
+    def strip_model(
+        self, model: TModel, do_copy: bool = False, strip_format: StripFormat = StripFormat.NATIVE
+    ) -> TModel:
         """
         Strips auxiliary layers that were used for the model compression, as it's
         only needed for training. The method is used before exporting the model
         in the target format.
 
         :param model: The compressed model.
         :param do_copy: Modify copy of the model, defaults to False.
+        :param strip format: Describes the format in which model is saved after strip.
         :return: The stripped model.
         """
         if do_copy:
@@ -256,16 +260,17 @@ def prepare_for_export(self) -> None:
         """
         self._model = self.strip_model(self._model)
 
-    def strip(self, do_copy: bool = True) -> TModel:  # type: ignore[type-var]
+    def strip(self, do_copy: bool = True, strip_format: StripFormat = StripFormat.NATIVE) -> TModel:  # type: ignore[type-var]
         """
-        Returns the model object with as much custom NNCF additions as possible removed
-        while still preserving the functioning of the model object as a compressed model.
+        Removes auxiliary layers and operations added during the compression process, resulting in a clean
+        model ready for deployment. The functionality of the model object is still preserved as a compressed model.
 
         :param do_copy: If True (default), will return a copy of the currently associated model object. If False,
           will return the currently associated model object "stripped" in-place.
+        :param strip format: Describes the format in which model is saved after strip.
         :return: The stripped model.
         """
-        return self.strip_model(self.model, do_copy)  # type: ignore
+        return self.strip_model(self.model, do_copy, strip_format)  # type: ignore
 
     @abstractmethod
     def export_model(

nncf/common/composite_compression.py

@@ -23,6 +23,7 @@
 from nncf.common.utils.backend import BackendType
 from nncf.common.utils.backend import copy_model
 from nncf.common.utils.backend import get_backend
+from nncf.parameters import StripFormat
 
 
 class CompositeCompressionLoss(CompressionLoss):
@@ -276,12 +277,12 @@ def prepare_for_export(self) -> None:
             stripped_model = ctrl.strip_model(stripped_model)
         self._model = stripped_model
 
-    def strip(self, do_copy: bool = True) -> TModel:  # type: ignore
+    def strip(self, do_copy: bool = True, strip_format: StripFormat = StripFormat.NATIVE) -> TModel:  # type: ignore
         model = self.model
         if do_copy:
             model = copy_model(model)
         for ctrl in self.child_ctrls:
-            model = ctrl.strip_model(model, do_copy=False)
+            model = ctrl.strip_model(model, do_copy=False, strip_format=strip_format)
         return model  # type: ignore
 
     @property

nncf/common/strip.py

@@ -16,6 +16,7 @@
 from nncf.common.utils.api_marker import api
 from nncf.common.utils.backend import BackendType
 from nncf.common.utils.backend import get_backend
+from nncf.parameters import StripFormat
 from nncf.telemetry.decorator import tracked_function
 from nncf.telemetry.events import MODEL_BASED_CATEGORY
 from nncf.telemetry.extractors import FunctionCallTelemetryExtractor
@@ -25,25 +26,26 @@
 
 @api(canonical_alias="nncf.strip")
 @tracked_function(category=MODEL_BASED_CATEGORY, extractors=[FunctionCallTelemetryExtractor("nncf.strip")])
-def strip(model: TModel, do_copy: bool = True) -> TModel:
+def strip(model: TModel, do_copy: bool = True, strip_format: StripFormat = StripFormat.NATIVE) -> TModel:
     """
-    Returns the model object with as much custom NNCF additions as possible removed
-    while still preserving the functioning of the model object as a compressed model.
+    Removes auxiliary layers and operations added during the compression process, resulting in a clean
+    model ready for deployment. The functionality of the model object is still preserved as a compressed model.
 
     :param model: The compressed model.
     :param do_copy: If True (default), will return a copy of the currently associated model object. If False,
       will return the currently associated model object "stripped" in-place.
+    :param strip format: Describes the format in which model is saved after strip.
     :return: The stripped model.
     """
     model_backend = get_backend(model)
     if model_backend == BackendType.TORCH:
         from nncf.torch.strip import strip as strip_pt
 
-        return strip_pt(model, do_copy)  # type: ignore
+        return strip_pt(model, do_copy, strip_format)  # type: ignore
     elif model_backend == BackendType.TENSORFLOW:
         from nncf.tensorflow.strip import strip as strip_tf
 
-        return strip_tf(model, do_copy)  # type: ignore
+        return strip_tf(model, do_copy, strip_format)  # type: ignore
 
-    msg = f"Method `strip` does not support for {model_backend.value} backend."
+    msg = f"Method `strip` does not support {model_backend.value} backend."
     raise nncf.UnsupportedBackendError(msg)

nncf/experimental/tensorflow/quantization/algorithm.py

@@ -35,6 +35,7 @@
 from nncf.experimental.tensorflow.quantization.init_range import RangeInitializerV2
 from nncf.experimental.tensorflow.quantization.init_range import TFRangeInitParamsV2
 from nncf.experimental.tensorflow.quantization.quantizers import create_quantizer
+from nncf.parameters import StripFormat
 from nncf.tensorflow.algorithm_selector import TF_COMPRESSION_ALGORITHMS
 from nncf.tensorflow.graph.metatypes.tf_ops import TFOpWithWeightsMetatype
 from nncf.tensorflow.graph.transformations.commands import TFInsertionCommand
@@ -353,7 +354,9 @@ def apply_to(self, model: NNCFNetwork) -> NNCFNetwork:
 
 
 class QuantizationControllerV2(QuantizationController):
-    def strip_model(self, model: NNCFNetwork, do_copy: bool = False) -> NNCFNetwork:
+    def strip_model(
+        self, model: NNCFNetwork, do_copy: bool = False, strip_format: StripFormat = StripFormat.NATIVE
+    ) -> NNCFNetwork:
         if do_copy:
             model = copy_model(model)
         return model

nncf/experimental/torch2/function_hook/graph/build_graph_mode.py

@@ -373,6 +373,6 @@ def build_graph(model: nn.Module, *args: Any, **kwargs: Any) -> nx.MultiDiGraph:
             with GraphBuilderMode(model=model, hook_storage=get_hook_storage(model)) as ctx:
                 args, kwargs = ctx.process_model_inputs(args, kwargs)
                 wrapped_forward = cast(ForwardWithHooks, model.forward)
-                outputs = wrapped_forward._func(*args, **kwargs)
+                outputs = wrapped_forward.orig_forward(*args, **kwargs)
                 outputs = ctx.process_model_outputs(outputs)
     return ctx.graph

nncf/experimental/torch2/function_hook/hook_executor_mode.py

@@ -41,6 +41,8 @@
     "size",
     "is_floating_point",
     "_set_grad_enabled",
+    "_parse_to",
+    "_has_compatible_shallow_copy_type",
 ]
 
 
@@ -517,3 +519,17 @@ def disable(self) -> Iterator[None]:
         self.enabled = False
         yield
         self.enabled = ret
+
+
+@contextmanager
+def disable_function_hook_mode() -> Iterator[None]:
+    """
+    Temporarily disables the function tracing and execution hooks within a context.
+    """
+    enabled_modes = torch.overrides._get_current_function_mode_stack()  # type: ignore[no-untyped-call]
+    state = {(mode, mode.enabled) for mode in enabled_modes if isinstance(mode, FunctionHookMode)}
+    for mode, _ in state:
+        mode.enabled = False
+    yield
+    for mode, enabled in state:
+        mode.enabled = enabled