Comments

.github/workflows/post_pr_merge.yml

@@ -16,6 +16,9 @@ on:
  - develop
  types:
  - closed
+ paths-ignore:
+ - '**/*.md'
+ - 'docs/**/*'
 
 jobs:
  upload-coverage-common:

.github/workflows/precommit.yml

@@ -95,50 +95,6 @@ jobs:
  token: ${{ secrets.CODECOV_TOKEN }}
  name: coverage_openvino
  flags: OPENVINO
- torchFX:
- timeout-minutes: 40
- defaults:
- run:
- shell: bash
- runs-on: ubuntu-20.04-8-cores
- env:
- DEBIAN_FRONTEND: noninteractive
- steps:
- - name: Install dependencies
- run : |
- sudo apt-get update
- sudo apt-get --assume-yes install gcc g++ build-essential ninja-build libgl1-mesa-dev libglib2.0-0
- - uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29 # v4.1.6
- with:
- lfs: true
- - uses: actions/setup-python@0a5c61591373683505ea898e09a3ea4f39ef2b9c # v5.0.0
- with:
- python-version: 3.8.18
- cache: pip
- - name: Runner info
- continue-on-error: true
- run: |
- cat /etc/*release
- cat /proc/cpuinfo
- - name: Install NNCF and test requirements
- run: make install-torch-fx-test
- - name: Run TorchFX precommit test scope
- run: |
- make test-torch-fx
- env:
- NNCF_COVERAGE: 1
- NUM_WORKERS: 4
- - name: Upload coverage report as artifact
- uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808 # v4.3.3
- with:
- name: coverage_fx_cpu
- path: ./coverage.xml
- - name: Upload coverage report to codecov
- uses: codecov/codecov-action@125fc84a9a348dbcf27191600683ec096ec9021c # v4.4.1
- with:
- token: ${{ secrets.CODECOV_TOKEN }}
- name: coverage_fx_cpu
- flags: TORCH
  pytorch-cpu:
  timeout-minutes: 40
  defaults:

Makefile

@@ -153,6 +153,7 @@ test-torch-cuda:
 
 test-torch-nightly:
  pytest ${COVERAGE_ARGS} tests/torch -m nightly --junitxml ${JUNITXML_PATH} $(DATA_ARG)
+ test-torch-fx
 
 test-torch-weekly:
  pytest ${COVERAGE_ARGS} tests/torch -m weekly \

nncf/common/factory.py

@@ -43,7 +43,7 @@ def create(model: TModel) -> NNCFGraph:
 
  return GraphConverter.create_nncf_graph(model)
  if model_backend == BackendType.TORCH_FX:
- from nncf.experimental.torch_fx.nncf_graph_builder import GraphConverter
+ from nncf.experimental.torch.fx.nncf_graph_builder import GraphConverter
 
  return GraphConverter.create_nncf_graph(model)
  if model_backend == BackendType.TORCH:
@@ -77,7 +77,7 @@ def create(model: TModel, inplace: bool = False) -> ModelTransformer:
 
  return PTModelTransformer(model)
  if model_backend == BackendType.TORCH_FX:
- from nncf.experimental.torch_fx.model_transformer import FXModelTransformer
+ from nncf.experimental.torch.fx.model_transformer import FXModelTransformer
 
  return FXModelTransformer(model)
  raise nncf.UnsupportedBackendError(
@@ -108,7 +108,7 @@ def create(model: TModel) -> Engine:
 
  return PTEngine(model)
  if model_backend == BackendType.TORCH_FX:
- from nncf.experimental.torch_fx.engine import FXEngine
+ from nncf.experimental.torch.fx.engine import FXEngine
 
  return FXEngine(model)
  raise nncf.UnsupportedBackendError(
@@ -164,7 +164,7 @@ def create(model: TModel, dataset: Dataset) -> aggregator.StatisticsAggregator:
 
  return PTStatisticsAggregator(dataset)
  if model_backend == BackendType.TORCH_FX:
- from nncf.experimental.torch_fx.statistics.aggregator import FXStatisticsAggregator
+ from nncf.experimental.torch.fx.statistics.aggregator import FXStatisticsAggregator
 
  return FXStatisticsAggregator(dataset)
  raise nncf.UnsupportedBackendError(

nncf/experimental/torch_fx/nncf_graph_builder.py → nncf/experimental/torch/fx/nncf_graph_builder.py

@@ -21,9 +21,9 @@
 from nncf.common.graph.layer_attributes import Dtype
 from nncf.common.graph.operator_metatypes import UnknownMetatype
 from nncf.common.logging import nncf_logger
-from nncf.experimental.torch_fx.transformations import separate_conv_and_bias
-from nncf.experimental.torch_fx.transformations import separate_linear_and_bias
-from nncf.experimental.torch_fx.transformations import view_to_reshape
+from nncf.experimental.torch.fx.transformations import separate_conv_and_bias
+from nncf.experimental.torch.fx.transformations import separate_linear_and_bias
+from nncf.experimental.torch.fx.transformations import view_to_reshape
 from nncf.torch.graph.graph import PTNNCFGraph
 from nncf.torch.graph.operator_metatypes import PT_OPERATOR_METATYPES
 

nncf/experimental/torch_fx/quantization/quantize_model.py → nncf/experimental/torch/fx/quantization/quantize_model.py

@@ -23,10 +23,11 @@
 
 import nncf
 from nncf.common.factory import NNCFGraphFactory
+from nncf.common.logging import nncf_logger
 from nncf.common.quantization.structs import QuantizationPreset
 from nncf.common.quantization.structs import QuantizationScheme
 from nncf.data import Dataset
-from nncf.experimental.torch_fx.transformations import merge_conv_and_bias
+from nncf.experimental.torch.fx.transformations import merge_conv_and_bias
 from nncf.parameters import ModelType
 from nncf.parameters import QuantizationMode
 from nncf.parameters import TargetDevice
@@ -53,6 +54,11 @@ def quantize_impl(
  """
  Implementation of the `quantize()` method for the Torch FX backend.
  """
+ nncf_logger.warning(
+ "Experimental Torch FX quantization backend is being used for the given torch.fx.GraphModule model."
+ " Torch FX PTQ is an experimental feature, consider using Torch or OpenVino PTQ backends"
+ " in case of errors or a poor model performance."
+ )
  if fast_bias_correction is False:
  raise ValueError(f"fast_bias_correction={fast_bias_correction} is not supported")
  if target_device == TargetDevice.CPU_SPR:

nncf/experimental/torch_fx/statistics/aggregator.py → nncf/experimental/torch/fx/statistics/aggregator.py

@@ -21,7 +21,7 @@
 from nncf.common.tensor_statistics.aggregator import StatisticPointsContainer
 from nncf.common.tensor_statistics.aggregator import StatisticsAggregator
 from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
-from nncf.experimental.torch_fx.model_transformer import FXModuleInsertionCommand
+from nncf.experimental.torch.fx.model_transformer import FXModuleInsertionCommand
 from nncf.tensor import Tensor
 from nncf.torch.nncf_network import NNCFNetwork
 from nncf.torch.return_types import maybe_get_values_from_torch_return_type

nncf/experimental/torch_fx/transformations.py → nncf/experimental/torch/fx/transformations.py

@@ -21,7 +21,7 @@
 
 from nncf.common.graph.graph import NNCFNode
 from nncf.common.graph.transformations.commands import TargetType
-from nncf.experimental.torch_fx.model_transformer import FXModelTransformer
+from nncf.experimental.torch.fx.model_transformer import FXModelTransformer
 from nncf.torch.graph.transformations.commands import PTTargetPoint
 
 

nncf/quantization/algorithms/fast_bias_correction/torch_fx_backend.py

@@ -22,8 +22,8 @@
 from nncf.common.graph.definitions import NNCFGraphNodeType
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
-from nncf.experimental.torch_fx.model_transformer import FXApplyTransformationCommand
-from nncf.experimental.torch_fx.transformations import bias_update_transformation_builder
+from nncf.experimental.torch.fx.model_transformer import FXApplyTransformationCommand
+from nncf.experimental.torch.fx.transformations import bias_update_transformation_builder
 from nncf.quantization.algorithms.fast_bias_correction.backend import FastBiasCorrectionAlgoBackend
 from nncf.tensor import Tensor
 from nncf.torch.graph.transformations.commands import PTModelExtractionCommand
@@ -83,7 +83,7 @@ def create_input_data(shape: Tuple[int], data: List[Tensor], input_name: str, ch
  @staticmethod
  def get_bias_value(node: NNCFNode, nncf_graph: NNCFGraph, model: torch.fx.GraphModule) -> Tensor:
  # TODO: make a node_name_vs_node map to speed up the process
- from nncf.experimental.torch_fx.model_transformer import FXModelTransformer
+ from nncf.experimental.torch.fx.model_transformer import FXModelTransformer
 
  bias_node = nncf_graph.get_next_nodes(node)[0]
  graph_bias_node = FXModelTransformer.get_graph_node_by_name(model.graph, bias_node.node_name)

nncf/quantization/algorithms/min_max/torch_fx_backend.py

@@ -27,8 +27,8 @@
 from nncf.experimental.common.tensor_statistics.collectors import AGGREGATORS_MAP
 from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
 from nncf.experimental.common.tensor_statistics.statistics import MinMaxTensorStatistic
-from nncf.experimental.torch_fx.model_transformer import FXApplyTransformationCommand
-from nncf.experimental.torch_fx.transformations import qdq_insertion_tranformation_builder
+from nncf.experimental.torch.fx.model_transformer import FXApplyTransformationCommand
+from nncf.experimental.torch.fx.transformations import qdq_insertion_tranformation_builder
 from nncf.parameters import ModelType
 from nncf.parameters import TargetDevice
 from nncf.quantization.advanced_parameters import StatisticsType

nncf/quantization/quantize_model.py

@@ -229,7 +229,7 @@ def quantize(
  advanced_parameters=advanced_parameters,
  )
  if backend == BackendType.TORCH_FX:
- from nncf.experimental.torch_fx.quantization.quantize_model import quantize_impl
+ from nncf.experimental.torch.fx.quantization.quantize_model import quantize_impl
 
  return quantize_impl(
  model=model,