[PDPD] Fix FastRCNN unsupported-operation crash on expand_2.tmp_0 (dynamic-rank Broadcast)

[liubo-intel]

Details:

• The Paddle expand_v2 translator can emit an opset Broadcast whose target-shape
  vector has a dynamic length, producing a dynamic-rank output. When such an
  expand_v2 sits inside a control-flow body (e.g. PaddleDetection faster_rcnn_r50_1x_coco),
  CPU plugin compilation will crash.
  this pr pins the Broadcast target-shape length to the statically-known expand output
  rank, keeping the Broadcast output rank static.

Tickets:

• CVS-190358
• CVS-141796

[liubo-intel]

new added testcase passed:

[Copilot]

Pull request overview

This PR addresses a Paddle frontend edge case where expand_v2 conversion can produce an OpenVINO Broadcast whose target-shape tensor has dynamic length, leading to a dynamic-rank Broadcast output and a CPU plugin compilation crash (notably in FastRCNN-like control-flow bodies).

Changes:

• In the Paddle expand_v2 translator, pin the Broadcast target-shape vector length to the statically known output rank (using Gather with constant indices).
• Add a new Paddle model generator case intended to exercise expand_v2 inside control flow.
• Register the new model in the Paddle frontend “fuzzy op” CPU-execution test set.

Reviewed changes

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

File	Description
src/frontends/paddle/src/op/expand_v2.cpp	Pins Broadcast target-shape length to static output rank via Gather to avoid dynamic-rank Broadcast outputs.
src/frontends/paddle/tests/test_models/gen_scripts/generate_expand_v2.py	Adds a control-flow model generator case for expand_v2 and reformats existing code.
src/frontends/paddle/tests/op_fuzzy.cpp	Includes the newly generated control-flow model in the CPU-backed fuzzy execution suite.

[yuxu42]

Hi @azhai219 could you review the PR?

[azhai219]

Hi @azhai219 could you review the PR?

ok

[github-actions]

CI Doctor — Merge Queue failure on this PR

Pipeline: Linux (Ubuntu 22.04, ARM64 cross-compilation, Python 3.11)
Failure: smoke_ConvertCPULayerTest - Value of: primTypeCheck(primType) is unexpected
Automatic restart: ❌ Not triggered — please retry the merge queue manually (the failure is unrelated to this PR's changes)

Possible remedy

1. Re-queue this PR in the merge queue — the failure is a pre-existing ARM64/ACL issue unrelated to the Paddle frontend changes in this PR.
2. If the failure recurs, report it to the CPU plugin team so they can investigate the ACL primitive type registration for i8/u8 conversion.
3. The CPU plugin team can add the specific test variant to the expected-failures list as a short-term workaround.

What happened

The ARM64 functional test smoke_ConvertCPULayerTest_4D_Dynamic (i8→u8, nchw, primitive=acl) failed at cpu_test_utils.cpp:248 with primType is unexpected : unknown_i8 Expected : ref_i8. This same test class has failed on at least 3 other unrelated PRs in the past 9 days (with f32, i32, and now i8 precisions), confirming it is a pre-existing flaky issue in the ARM64 ACL backend — not caused by this PR.

Generated by CI Failure Doctor — Merge Queue · 135.3 AIC · ⌖ 6.47 AIC · ⊞ 20K · ◷