[GPU] Prevent unsafe in-place crop optimization for 2D-to-1D bbox splits

[zhanmyz]

Details:

• Prevent unsafe runtime in-place crop optimization for 2D-to-1D rank-reducing reshape patterns.
• Add a regression unit test for the Faster RCNN / SSD bbox-coordinate pattern.

Description of the issue (symptom, root cause, and resolution)

• Symptom:
  After commit 3f24992895e41aa669b3743c8b6a7f499f3d789a, several GPU object-detection models produced incorrect detection results. The original reproducer was:

  • TF_V2_Faster_RCNN_Inception_ResNet_v2_atrous_coco
  • Expected detections: 5
  • GPU inference detections: 25
  • Failing output layer: tf_detections

  The same regression pattern was also observed for:

  • TF_V2_Faster_RCNN_ResNet50_v1_atrous_coco
  • TF_V2_Mask_RCNN_ResNetv2_atrous_coco
  • ONNX_Runtime_ssd_mobilenet_V1_coco_mlperf_opset10

• Root Cause
  Commit 3f24992895e41aa669b3743c8b6a7f499f3d789a enabled runtime padding propagation for the TransposeSplitMatcher QKV pattern. The new condition in reshape_node::is_runtime_propagatable_padding() accepted any axis-1 crop with a size-1 dimension followed by a rank-reducing reshape:

  axis == 1 && input_pshape[1] == 1 && output_rank + 1 == input_rank

  This condition was intended for the QWen-VL pattern:

  [-1, 3, H, S] -> [-1, 1, H, S] -> [-1, H, S]
           4D           4D              3D
  

  However, it also accepted the object-detection bbox-coordinate pattern:

  [N, 4] -> [N, 1] -> [N]
     2D       2D      1D
  

  The latter is not safe for runtime padding propagation. In the affected models, Split([N, 4]) produces four coordinate crops [N, 1], which are squeezed to one-dimensional vectors [N] and consumed by bbox decode arithmetic. The in-place crop optimization exposed a strided/padded view to downstream kernels that expected a contiguous one-dimensional vector. This corrupted decoded box coordinates and caused NMS to return an incorrect number of detections.
  The existing unit tests added by commit 3f24992895e covered the valid QKV/TransposeSplitMatcher path and other padding propagation cases, but did not cover the invalid 2D-to-1D bbox-coordinate boundary case.

• Resolution:
  Restrict the axis-1 rank-reducing padding propagation rule to outputs with rank at least 2:

  if (prim->output_partial_shape.size() >= 2 &&
      prim->output_partial_shape.size() + 1 == input_pshape.size()) {
      return true;
  }

  The same guard is applied in the vl_sdpa path so both build-time and runtime decisions use the same safety contract.
  This preserves the intended QWen-VL optimization (4D -> 3D) while rejecting the unsafe bbox-coordinate case (2D -> 1D). The change is generic and is not tied to a specific model.

Performance and compatibility

• The fix only disables the newly introduced padding optimization for unsafe 2D-to-1D rank reduction. This path was not optimized before commit 3f24992895e, so the change restores the previous behavior for these models and avoids an accuracy regression. The valid QWen-VL 4D-to-3D optimization remains enabled.

The code and line that caused this issue (if it is not changed directly)

openvino/src/plugins/intel_gpu/src/graph/include/reshape_inst.h

Lines 137 to 141 in 1872889

	if (axis == 1 && !input_pshape[1].is_dynamic() && input_pshape[1].get_length() == 1) {
	if (prim->output_partial_shape.size() + 1 == input_pshape.size()) {
	return true;
	}
	}

Reproduction step

python -m pytest test_skip_ir_generation.py \
    --tb=native \
    --env_conf=.automation/env_config.yml \
    --test_conf=.automation/test_configs/desktop_test_config.yml \
    -m "not launch_only_if_manually_specified " \
    --pregen_irs=irs_mapping.csv \
    --tf_models_version=2 \
    --modules pipelines/production/tf_2x/heavy \
    -k "TF_V2_Faster_RCNN_Inception_ResNet_v2_atrous_coco_batch_1_device_GPU_precision_FP32" \
    --dynamism_type=None \
    --skip_ir_generation \
    --log-cli-level INFO 

Checklist

• Is it a proper fix? (not a workaround)
• Did you include test case for this fix, if necessary?
• Did you review existing test that can be extended to cover this scenario? Which test did you review?

Tickets:

• CVS-192143, CVS-192144, CVS-192235

[ceciliapeng2011]

@zhanmyz The PR LGTM. Please open it for official review and merge.

[zhanmyz]

@zhanmyz The PR LGTM. Please open it for official review and merge.

Sure, Added unit test to cover the changed code.

[e-ddykim]

Could you please check if the issue still exists in the latest master? I want to check if the issue was resolved by #37267

[zhanmyz]

Could you please check if the issue still exists in the latest master? I want to check if the issue was resolved by #37267

I’ve tested with the latest OV PR #37160 as well as PR #37267. Unfortunately, the issue still persists.
So it is not resolved by #37267.