[NNCFGraph] Migrate from nx.DiGraph to nx.MultiDiGraph

[daniil-lyakhov]

Changes

• NNCFGraph + InsertionPointGraph + PatternMatchingGraph are inhereted from nx.MultiDiGraph instead of nx.DiGraph
• SDPA Ignored pattern is updated with split operation as a branching metatype for OpenVINO/ONNX backends
• Hardware fusing pattern conv -> arithmetic is updated to have 1 and 2 degrees edges like on this picture:

• ROPE patern is updated to have several parallel input edges to the concat node

Reason for changes

• To make it possible to represent and compress models with Multi edges in NNCF
• To quantize YOLO26 SDPA block properly in OpenVINO/ONNX backends
• To quantize models like a = conv(x); return aa correctly (not quantizing the xx part)
• To apply the ROPE pattern correctly, as the original pattern has parallel edges

Related tickets

174691
179960

Tests

• MultiDiGraph NNCFGraph is tested here https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-1f470020da6ca32a3f1881ec05b5af9ac828bfd8d892ae1306c860553b960d5bR113-R175
• InsertionPointGraph and PatternMatchingGraph (+SDPA updated ignored pattern +conv+arithmetic HW fused pattern) are tested here https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-7cac922c198eebec6543d9da25eea4fb03aa6e352c4416cf157a41d70010d5feR522-R611
• ONNX MultiDiGraph building test: https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-577b44dd17279c195cfd652ef121c720c5a4278dbcd65cd407932cb7e9d2adccR148-R167
• OpenVINO MultiDiGraph bulding and quantization test: https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-b7a697bf639687bc47899b5166a4e42636c00162cd146739a931382fa1df8a34R1383-R1403
• Torch MultiDIGraph bulding and quantization test: https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-09664d18a9fe689dca4a78998cc8e702526d81b5685c1164b56ce2dcb74b139aR67-R78
• TorchFX MultiDiGraph building test https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-741156c097f1cd42544f5ba4a7dc31024a454b2f78581ebcd2e65f11377fa1a7R80
  (Could not quantize yolo26 SDPA subgraph correctly due to the issue [TorchFX] Do not propagate quantizers through __get_item__ when input/output has different numel #3109)
• Test ROPE pattern with 2 parallel edges: https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-32bc7464d48f05ead87262b533f2e0ce9ea2db101e9e8ae59cb64c65d6aba351R448-R450
• ONNX Synthetic parallel edges model added to check parallel edges nncf graph building: https://github.com/openvinotoolkit/nncf/pull/3843/changes#diff-9aec6e5f328ddb6b440894cb95dcc3b774d34bb67547e9aad6b065bd678c585fR248-R292

Jobs

Test WC: https://github.com/openvinotoolkit/nncf/actions/runs/21724379639 - Green
Test examples: https://github.com/openvinotoolkit/nncf/actions/runs/21724373555 - Green
PTQ: NNCF/job/manual/job/post_training_quantization/804/ - Green

[Copilot]

Pull request overview

This PR migrates NNCF graph representations from nx.DiGraph to nx.MultiDiGraph to support parallel/multi-edges, and updates quantization/hardware-fusing/ignored-pattern logic (notably SDPA and RoPE) plus corresponding tests and reference graphs.

Changes:

• Switched core graph utilities, pattern matching, and insertion-point graphs to nx.MultiDiGraph and adapted edge handling APIs (get_edges, keyed edges in .dot files).
• Updated HW fused patterns (conv -> arithmetic) and ignored patterns (RoPE + SDPA) to account for parallel edges/branching.
• Added/updated synthetic models, test fixtures, and many .dot/JSON references to validate multi-edge behavior across Torch/FX, OpenVINO, and ONNX.

Reviewed changes

Copilot reviewed 97 out of 114 changed files in this pull request and generated 13 comments.

Show a summary per file

File	Description
tools/render_dot_to_svg.py	Uses nx.MultiDiGraph when rendering .dot graphs to preserve multi-edges.
tools/clip_dot.py	Uses nx.MultiDiGraph for reading and producing clipped graphs with possible parallel edges.
tests/torch/utils.py	Updates comparable graph conversion to output nx.MultiDiGraph.
tests/torch/test_graph_analysis.py	Updates mock graph to MultiDiGraph and edge construction without parallel-port field.
tests/torch/fx/test_weights_compression.py	Updates edge retrieval for multi-edge API and parameterizes RoPE degree.
tests/torch/fx/test_quantizer_config.py	Adds fixtures for degree-2 arithmetic patterns and split-based transformer graphs.
tests/torch/fx/test_quantizer.py	Removes parallel_input_port_ids usage in normalization helper.
tests/torch/fx/test_models.py	Adds YOLO26 attention block to FX model cases.
tests/torch/function_hook/quantization/test_weights_compression.py	Parameterizes RoPE model creation with degree.
tests/torch/function_hook/quantization/test_quantizer_config.py	Adds fixtures mirroring FX config tests (degree-2 + split transformer).
tests/torch/function_hook/quantization/test_quantized_graphs.py	Adds models for parallel edges + YOLO26; updates RoPE to degree=2.
tests/torch/function_hook/quantization/helper.py	Adds builder for degree-2 conv->arithmetic graph template.
tests/torch/data/fx/yolo26_attn_block.dot	Adds FX reference .dot for YOLO26 attention block.
tests/torch/data/fx/reference_metatypes/yolo26_attn_block.json	Adds metatype reference for YOLO26 attention block.
tests/torch/data/function_hook/sparsify_activations/three_linear_sparse_activations.dot	Updates .dot format to include multi-edge key= and non-strict header.
tests/torch/data/function_hook/sparsify_activations/three_linear_int8_sym_weights_sparse_activations.dot	Same .dot updates for int8 sparse-activations reference.
tests/torch/data/function_hook/sparsify_activations/three_linear_ignore1_sparse_activations.dot	Same .dot updates for ignore-scope reference.
tests/torch/data/function_hook/sparsify_activations/three_linear_ignore1_int8_sym_weights_sparse_activations.dot	Same .dot updates for ignore-scope + int8 reference.
tests/torch/data/function_hook/sparsify_activations/linear_sparse_activations.dot	Same .dot updates for single-linear sparse reference.
tests/torch/data/function_hook/sparsify_activations/linear_int8_sym_weights_sparse_activations.dot	Same .dot updates for single-linear + int8 reference.
tests/torch/data/function_hook/quantization/test_quantized_graphs/yolo26_attn_block.dot	Adds quantized-graph .dot reference for YOLO26 attention block.
tests/torch/data/function_hook/quantization/test_quantized_graphs/unbind_scaled_dot_product_attention_model.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/quantization/test_quantized_graphs/shared_model.dot	Replaces parallel_input_port_ids with explicit parallel edges via key.
tests/torch/data/function_hook/quantization/test_quantized_graphs/scaled_dot_product_attention_model.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/quantization/test_quantized_graphs/rope_model.dot	Updates RoPE .dot to represent parallel concat inputs as multi-edges.
tests/torch/data/function_hook/quantization/test_quantized_graphs/parallel_edges_model.dot	Adds .dot reference for explicit parallel-edge model.
tests/torch/data/function_hook/quantization/test_quantized_graphs/lenet.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/quantization/test_quantized_graphs/embedding_model.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/pruning_and_quantization/prune_ptq_model.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/nncf_graph/model_graph_with_shared_parameters.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/nncf_graph/model_graph_gru.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/nncf_graph/convert_to_nncf_graph_multi_edges.dot	Updates reference from “parallel ports” to explicit multi-edges.
tests/torch/data/function_hook/nncf_graph/convert_to_nncf_graph.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/compress_weights/fq_lora/shared_weights_all_layers_True.dot	Updates .dot to keyed multi-edges.
tests/torch/data/function_hook/compress_weights/fq_lora/shared_weights_all_layers_False.dot	Updates .dot to keyed multi-edges.
tests/openvino/native/test_nncf_graph_builder.py	Removes parallel_input_port_ids from edge fixtures.
tests/openvino/native/test_model_utils.py	Updates direct edge-attr access to include MultiDiGraph edge key.
tests/openvino/native/quantization/test_weights_compression.py	Parameterizes RoPE model creation with degree.
tests/openvino/native/quantization/test_quantizer_config.py	Adds split + degree-2 graph fixtures and includes OVSplitMetatype.
tests/openvino/native/quantization/test_graphs.py	Updates transformer model paramization to use partial(RoPEModel, degree=2).
tests/openvino/native/models.py	Updates RoPE model to represent concat degree and adds YOLO26 attention block model.
tests/openvino/native/data/2025.4/reference_scales/YOLO26AttentionBlock_performance.json	Adds reference scales for YOLO26 attention block.
tests/openvino/native/data/2025.4/reference_scales/YOLO26AttentionBlock_mixed.json	Adds mixed reference scales for YOLO26 attention block.
tests/openvino/native/data/2025.4/reference_graphs/quantized/YOLO26AttentionBlock.dot	Adds quantized reference .dot for YOLO26 attention block with keyed edges.
tests/openvino/native/data/2025.4/reference_graphs/original_nncf_graph/YOLO26AttentionBlock.dot	Adds original-graph reference .dot for YOLO26 attention block with keyed edges.
tests/openvino/native/data/2024.4/reference_graphs/quantized/RoPEModel.dot	Updates RoPE reference .dot to keyed multi-edges and renamed nodes.
tests/onnx/test_nncf_graph_builder.py	Adds ONNX multi-edge conversion test using Split.
tests/onnx/quantization/test_weights_compression.py	Parameterizes RoPE concat degree for ONNX model builder.
tests/onnx/quantization/test_quantizer_config.py	Adds split + degree-2 fixtures and includes ONNXSplitMetatype.
tests/onnx/models.py	Adds synthetic ONNX model with parallel edges via repeated Concat inputs.
tests/onnx/data/reference_graphs/quantization/synthetic/multi_input_output_parallel_edges_model.dot	Adds keyed-edge reference graph for ONNX parallel-edges synthetic model.
tests/onnx/data/reference_graphs/original_nncf_graph/synthetic/multi_input_output_parallel_edges_model.dot	Adds keyed-edge original-graph reference for ONNX parallel-edges model.
tests/onnx/common.py	Adds ModelBuilder.add_split helper for Split-based models.
tests/cross_fw/test_templates/test_quantizer_config.py	Adds template tests/states for degree-2 conv+arith and split-transformer patterns.
tests/cross_fw/test_templates/template_test_weights_compression.py	Parameterizes RoPE compression test by concat degree.
tests/cross_fw/test_templates/models.py	Adds graph templates for arithmetic degree-2 and split-transformer.
tests/cross_fw/test_templates/helpers.py	Updates torch RoPE model to accept degree; adds ParallelEdgesModel + YOLO26AttentionBlock.
tests/cross_fw/shared/nx_graph.py	Updates DOT sorting + graph comparison to MultiDiGraph and keyed edges.
tests/common/quantization/test_quantizer_propagation_solver.py	Converts mock graphs to MultiDiGraph and updates edge indexing to include keys.
tests/common/quantization/test_quantizer_propagation_graph.py	Converts mock graphs to MultiDiGraph and updates edge indexing to include keys.
tests/common/quantization/mock_graphs.py	Converts mock graph utilities to MultiDiGraph; ensures in_edges/out_edges use keys.
tests/common/graph/test_nncf_graph.py	Adds explicit multi-edge tests and duplicate-edge validation test.
tests/common/graph/test_dot_file_rw.py	Updates DOT read/write test graph type to MultiDiGraph.
tests/common/data/reference_graphs/dot_rw_reference.dot	Updates DOT reference to non-strict digraph and keyed edge format.
src/nncf/torch/quantization/ignored_patterns.py	Reuses shared RoPE pattern builder to support parallel edges.
src/nncf/torch/model_graph_manager.py	Updates constant-input resolution for multi-edge graph API.
src/nncf/torch/hardware/fused_patterns.py	Updates HW fused patterns to include parallel-edge degree alternatives.
src/nncf/torch/graph/graph.py	Updates insertion-point shape extraction using edge-by-port APIs.
src/nncf/torch/function_hook/nncf_graph/nncf_graph_builder.py	Emits explicit edges for each meta edge instead of parallel-port encoding.
src/nncf/quantization/passes.py	Updates node-removal reconnect logic to handle multi-edges.
src/nncf/quantization/ignored_patterns.py	Introduces shared create_rope_pattern supporting concat degree alternatives.
src/nncf/quantization/algorithms/weight_compression/torch_fx_backend.py	Updates weight-shape/axes helpers to handle multi-edge APIs and validates edge uniqueness.
src/nncf/quantization/algorithms/weight_compression/torch_backend.py	Updates weight/activation port resolution to iterate over multiple edges.
src/nncf/quantization/algorithms/smooth_quant/algorithm.py	Updates grouping logic to validate and use a single edge between nodes.
src/nncf/openvino/quantization/ignored_patterns.py	Reuses shared RoPE pattern builder and updates SDPA ignored pattern to include split.
src/nncf/openvino/hardware/fused_patterns.py	Updates HW fused patterns to include parallel-edge degree alternatives.
src/nncf/openvino/graph/nncf_graph_builder.py	Emits explicit edges per OpenVINO input port instead of parallel-port encoding.
src/nncf/onnx/quantization/ignored_patterns.py	Reuses shared RoPE pattern builder and updates SDPA ignored pattern to include split.
src/nncf/onnx/hardware/fused_patterns.py	Updates HW fused patterns to include parallel-edge degree alternatives.
src/nncf/onnx/graph/onnx_helper.py	Updates input-port lookup to return all matching input ports for parallel edges.
src/nncf/onnx/graph/nncf_graph_builder.py	Updates ONNX graph builder to create explicit multi-edges and disambiguate repeated inputs.
src/nncf/experimental/torch/sparsify_activations/torch_backend.py	Updates activation-port resolution to iterate over multiple edges.
src/nncf/common/utils/dot_file_rw.py	Updates DOT reader to return nx.MultiDiGraph.
src/nncf/common/insertion_point_graph.py	Migrates insertion-point graph to MultiDiGraph and updates edge/key handling.
src/nncf/common/graph/patterns/patterns.py	Migrates GraphPattern to MultiDiGraph and adds join_patterns_parallel.
src/nncf/common/graph/graph_matching.py	Uses MultiDiGraphMatcher for pattern matching on MultiDiGraph graphs.
src/nncf/common/graph/graph.py	Migrates core NNCFGraph to MultiDiGraph; replaces parallel-port encoding with explicit multi-edges.

Comments suppressed due to low confidence (1)

src/nncf/common/utils/dot_file_rw.py:1

• read_dot_graph now returns nx.MultiDiGraph, but write_dot_graph still type-annotates nx.DiGraph. Consider widening the annotation (e.g., nx.MultiDiGraph or a union of graph types) to reflect actual usage and prevent confusion in downstream call sites.

# Copyright (c) 2026 Intel Corporation

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.