[HETERO] Fix multi-hop cycle in SubgraphCollector

[wgzintel]

Details

Fixes OPENVINO_ASSERT("Cannot sort subgraphs!") in ov::hetero::SubgraphCollector::run() on real workloads (reproduced on yolo26seg with HETERO:GPU,CPU, 736 ops / 643 → GPU / 93 → CPU).

Root cause

The existing per-node cyclic-dependency check in split_cyclic_dependencies() (the bit-scan Phase 4a/4b) only catches cycles whose re-entry consumer is itself a node whose cyc_dep bitset is non-empty (same-subgraph data flowing back through one foreign subgraph into that node's inputs).

Two classes of subgraph-DAG cycles fall outside its scope, and both occur naturally in valid ov::Models:

1. Multi-hop subgraph-DAG cycles — sg_A → sg_B → sg_C → sg_D → sg_A. Producer and re-entry consumer are several subgraphs apart; no single node sees its own sg on the cycle.
2. Shared-graph-input cycles — a Constant (or other graph input) fans out to multiple consumers that Union-Find fuses into one subgraph; that fused subgraph then both produces and consumes data on the same neighbor subgraph. The cut edge is an input of the foreign-sg node, not of the same-sg node whose cyc_dep is non-empty, so the existing Phase 4b cannot promote it by construction.

Neither is a real data cycle in the ov::Model (which is always a DAG); both are artifacts of subgraph fusion that the topological sort in run() cannot resolve, leading to the assert.

Fix

Adds a subgraph-DAG SCC fallback that runs after the existing per-node heuristic. Each iteration:

1. build_subgraph_adjacency — derives the cross-subgraph DAG from _subgraph_inputs. Parallel edges between the same subgraph pair are de-duplicated; self-edges (producer_sg == owner_sg) are filtered.
2. find_non_trivial_scc_members — runs iterative Tarjan and returns subgraphs that belong to any SCC of size > 1. An exact SCC algorithm is required: a two-peel (forward + reverse Kahn) approximation also flags acyclic bridges between disjoint cycles, which would either waste promotion on acyclic subgraphs or trip the "no internal edge" assert.
3. isolate_one_scc_node — picks one SCC-member node with the fewest same-sg non-boundary inputs (tie-break: smaller topo index) and promotes all its same-sg inputs into _subgraph_inputs. Single-edge cuts diverge in practice — the chosen node re-merges via its other same-sg inputs in the next collect_subgraphs_ids round; full-dissolution explodes the partition. The "isolate one node" cut is the minimum that strictly reduces the non-singleton mass of SCC members each iteration.

Convergence is bounded by total node-input edges; two strict-progress asserts inside the loop catch any future strategy regression that fails to make progress:

• OPENVINO_ASSERT(promoted > 0, ...) — helper invariant.
• OPENVINO_ASSERT(_subgraph_inputs.size() > inputs_before_step, ...) — defensive against insert() returning all-duplicates.

Constant-sourced edges are eligible promotion targets (no is_graph_input_node filter on candidates). Without this, the SCC fallback would trip "no internal edge to promote" on shared-Constant cycles.

The rationale and convergence argument are documented inline above isolate_one_scc_node. The fallback's header comment explicitly classifies cases (a) and (b) above as first-class in-scope cases rather than exceptional safety-net cases.

Tests

New parametric cases under INSTANTIATE_TEST_SUITE_P(SubgraphCollectorParamTest, ...):

• split_cyclic_dependency_diamond — minimal diamond cycle requiring the existing per-node pass.
• split_cyclic_dependency_nested — stacked cycles, multi-iteration convergence of the existing pass.
• shared_const_indirect_bridge_cycle, shared_const_cross_device_fanout_cycle, shared_const_as_cycle_source — three shared-Constant cycle shapes the SCC fallback must resolve.
• shared_const_no_cycle_no_split, shared_const_indirect_bridge_no_cycle — negative cases (shared-Constant topology without an actual cycle must not be split).
• multi_hop_subgraph_scc_cycle — minimal synthesis of the multi-hop SCC observed on yolo26seg. expected_subgraph_count is std::optional<size_t> set to std::nullopt here because the exact partition is a property of the promotion strategy, not the correctness contract; the assertion under test is "run() does not assert Cannot sort subgraphs!" and "merge round-trip succeeds".

New standalone regression tests:

• SubgraphCollectorBridgeBetweenCyclesTest.bridge_subgraph_not_split — pins the "exact SCC, not two-peel approximation" invariant; an acyclic bridge between two disjoint cycles must stay un-split.
• SubgraphCollectorSharedConstSccTest.scc_with_only_constant_sourced_edges_converges — pins the Constant-source-filter removal: when every promotable candidate has a Constant producer, the SCC fallback must still converge instead of asserting "no internal edge to promote".

SubgraphCollectorTestParam::expected_subgraph_count was changed to std::optional<size_t> so opting out of an exact submodel-count assertion is explicit at the test definition site.

Validation

End-to-end on yolo26seg, HETERO:GPU,CPU (736 ops, 643 → GPU, 93 → CPU): SCC fallback converges in 145 iterations into 151 submodels; all submodels compile successfully and compile_model returns OK. Existing hetero unit and functional tests pass.

Impact

No public API or ABI change. Hetero plugin only; no change to bindings, frontends, or other plugins.

Tickets

• CVS-186543

AI Assistance:

• AI assistance used: yes

[Copilot]

Pull request overview

Fixes a Cannot sort subgraphs! assert in HETERO::compile_model when affinity assignment plus Union-Find through a shared graph-input boundary (e.g., a shared Constant) creates a multi-hop cycle in the subgraph DAG that the existing per-node cycle splitter cannot detect. Adds a subgraph-DAG SCC fallback after the per-node fix-point loop in SubgraphCollector::split_cyclic_dependencies().

Changes:

• Add a Kahn-based SCC detection pass over the subgraph DAG; for any cyclic subgraph, promote one same-subgraph non-boundary edge into _subgraph_inputs and re-run Union-Find until the DAG is acyclic.
• Add a regression test multi_hop_subgraph_scc_cycle synthesizing the yolo26s-seg 4-subgraph cycle.
• Allow expected_subgraph_count == 0 to disable exact subgraph-count assertion in the parameterized harness (partition shape is an implementation detail of the fallback).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/plugins/hetero/src/subgraph_collector.cpp	Adds <queue> include and a bounded SCC fallback loop that builds the cross-subgraph DAG, peels acyclic subgraphs via Kahn, and promotes one internal edge per iteration until acyclic.
src/plugins/hetero/tests/unit/subgraph_collector.cpp	New create_multi_hop_scc_cycle_model + parameterized test entry; loosens the exact expected_subgraph_count assertion when set to 0.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[WeldonWangwang]

Just want to know will there be a significant change in performance after adding subgraph-level SCC fallback?