YOLO11 / YOLO26 accuracy-aware quantization fails

[piotrgrubicki]

1. Train a YOLO26-S-Seg model
2. Once trained, Start quantization
3. Choose 2.7 as "Max accuracy drop (%)" and 12 as "Max number of iterations"
4. Start job

The job fails with the exception visible below.
NB: When executed using the default parameters, the job finishes successfully.

Job logs

{"text": "2026-08-07 11:40:27.672 | INFO     | logging:handle:1027 - Validation of initial model was started\n", "record": {"elapsed": {"repr": "0:01:59.802046", "seconds": 119.802046}, "exception": null, "extra": {}, "file": {"name": "__init__.py", "path": "logging\\__init__.py"}, "function": "handle", "level": {"icon": "ℹ️", "name": "INFO", "no": 20}, "line": 1027, "message": "Validation of initial model was started", "module": "__init__", "name": "logging", "process": {"id": 19032, "name": "job-quantize-506b7035-47ba-4d34-a1d9-fae6bc885cbd"}, "thread": {"id": 16652, "name": "MainThread"}, "time": {"repr": "2026-08-07 11:40:27.672509+02:00", "timestamp": 1786095627.672509}}}
{"text": "2026-08-07 11:40:28.390 | INFO     | logging:handle:1027 - Elapsed Time: 00:00:00\n", "record": {"elapsed": {"repr": "0:02:00.520154", "seconds": 120.520154}, "exception": null, "extra": {}, "file": {"name": "__init__.py", "path": "logging\\__init__.py"}, "function": "handle", "level": {"icon": "ℹ️", "name": "INFO", "no": 20}, "line": 1027, "message": "Elapsed Time: 00:00:00", "module": "__init__", "name": "logging", "process": {"id": 19032, "name": "job-quantize-506b7035-47ba-4d34-a1d9-fae6bc885cbd"}, "thread": {"id": 16652, "name": "MainThread"}, "time": {"repr": "2026-08-07 11:40:28.390617+02:00", "timestamp": 1786095628.390617}}}
{"text": "2026-08-07 11:41:12.175 | ERROR    | app.execution.base:_report_progress:122 - Failed: Run Quantization\nTraceback (most recent call last):\n\n  File \"main.py\", line 19, in <module>\n\n  File \"pyi_rth_multiprocessing.py\", line 48, in _freeze_support\n\n  File \"multiprocessing\\spawn.py\", line 122, in spawn_main\n\n  File \"multiprocessing\\spawn.py\", line 135, in _main\n\n  File \"multiprocessing\\process.py\", line 313, in _bootstrap\n\n  File \"multiprocessing\\process.py\", line 108, in run\n\n  File \"app\\core\\jobs\\exec\\process_run.py\", line 149, in _entrypoint\n\n  File \"app\\execution\\base.py\", line 92, in run\n\n  File \"app\\execution\\quantization\\getitune_quantizer.py\", line 344, in execute\n\n> File \"app\\execution\\base.py\", line 41, in wrapper\n\n  File \"app\\execution\\quantization\\getitune_quantizer.py\", line 257, in run_quantization\n\n  File \"getitune\\backend\\openvino\\engine.py\", line 476, in optimize\n    return model.optimize(\n           │     └ <function OVModel.optimize at 0x000001DE671758A0>\n           └ <getitune.backend.openvino.models.instance_segmentation.OVInstanceSegmentationModel object at 0x000001DE67461D30>\n\n  File \"getitune\\backend\\openvino\\models\\base.py\", line 376, in optimize\n    compressed_model = nncf.quantize_with_accuracy_control(\n                       │    └ <function quantize_with_accuracy_control at 0x000001DE4F057380>\n                       └ <module 'nncf' from 'C:\\\\Program Files\\\\WindowsApps\\\\intel.geti_3.1.0.0_neutral__aqcv7fbxf4wm4\\\\_internal\\\\nncf\\\\__init__.py'>\n\n  File \"nncf\\telemetry\\decorator.py\", line 81, in wrapped\n    retval = fn(*args, **kwargs)\n             │   │       └ {'model': <Model: 'Model0'\n             │   │         inputs[\n             │   │         <ConstOutput: names[x.1] shape[1,3,640,640] type: f32>\n             │   │         ]\n             │   │         outputs[\n             │   │         <ConstOutput: names[] sh...\n             │   └ ()\n             └ <function quantize_with_accuracy_control at 0x000001DE4F057240>\n\n  File \"nncf\\quantization\\quantize_model.py\", line 356, in quantize_with_accuracy_control\n    return quantize_with_accuracy_control_impl(  # type: ignore[no-any-return]\n           └ <function quantize_with_accuracy_control_impl at 0x000001DE67583380>\n\n  File \"nncf\\openvino\\quantization\\quantize_model.py\", line 240, in quantize_with_accuracy_control_impl\n    initial_metric_results = evaluator.collect_metric_results(model, validation_dataset, model_name=\"initial\")\n                             │         │                      │      └ <nncf.data.dataset.Dataset object at 0x000001DE67481BD0>\n                             │         │                      └ <Model: 'Model0'\n                             │         │                        inputs[\n                             │         │                        <ConstOutput: names[x.1] shape[1,3,640,640] type: f32>\n                             │         │                        ]\n                             │         │                        outputs[\n                             │         │                        <ConstOutput: names[] shape[1,39,8...\n                             │         └ <function Evaluator.collect_metric_results at 0x000001DE372DB600>\n                             └ <nncf.quantization.algorithms.accuracy_control.evaluator.Evaluator object at 0x000001DE677EFE00>\n\n  File \"nncf\\quantization\\algorithms\\accuracy_control\\evaluator.py\", line 331, in collect_metric_results\n    metric, values_for_each_item = self.validate_prepared_model(prepared_model, dataset)\n                                   │    │                       │               └ <nncf.data.dataset.Dataset object at 0x000001DE67481BD0>\n                                   │    │                       └ <nncf.quantization.algorithms.accuracy_control.openvino_backend.OVPreparedModel object at 0x000001DE677EFCB0>\n                                   │    └ <function Evaluator.validate_prepared_model at 0x000001DE372DB2E0>\n                                   └ <nncf.quantization.algorithms.accuracy_control.evaluator.Evaluator object at 0x000001DE677EFE00>\n\n  File \"nncf\\quantization\\algorithms\\accuracy_control\\evaluator.py\", line 164, in validate_prepared_model\n    metric, values_for_each_item = self._validation_fn(prepared_model.model_for_inference, validation_dataset)\n                                   │    │              │              │                    └ <nncf.quantization.algorithms.accuracy_control.evaluator.IterationCounter object at 0x000001DE67460D70>\n                                   │    │              │              └ <property object at 0x000001DE733F3F10>\n                                   │    │              └ <nncf.quantization.algorithms.accuracy_control.openvino_backend.OVPreparedModel object at 0x000001DE677EFCB0>\n                                   │    └ <function wrap_validation_fn.<locals>.wrapper at 0x000001DE3661EC00>\n                                   └ <nncf.quantization.algorithms.accuracy_control.evaluator.Evaluator object at 0x000001DE677EFE00>\n\n  File \"nncf\\quantization\\quantize_model.py\", line 261, in wrapper\n    retval = validation_fn(*args, **kwargs)\n             │              │       └ {}\n             │              └ (<CompiledModel:\n             │                inputs[\n             │                <ConstOutput: names[x.1] shape[1,3,640,640] type: f32>\n             │                ]\n             │                outputs[\n             │                <ConstOutput: names[] shape[1,39,8...\n             └ <function OVModel._create_validation_fn.<locals>.validation_fn at 0x000001DE672F80E0>\n\n  File \"getitune\\backend\\openvino\\models\\base.py\", line 462, in validation_fn\n    preds = _infer_compiled_model(compiled_model, data_batch)\n            │                     │               └ SampleBatch(images=tensor([[[[0., 0., 0.,  ..., 0., 0., 0.],\n            │                     │                           [0., 0., 0.,  ..., 0., 0., 0.],\n            │                     │                           [0., 0., 0.,...\n            │                     └ <CompiledModel:\n            │                       inputs[\n            │                       <ConstOutput: names[x.1] shape[1,3,640,640] type: f32>\n            │                       ]\n            │                       outputs[\n            │                       <ConstOutput: names[] shape[1,39,84...\n            └ <function OVModel._create_validation_fn.<locals>._infer_compiled_model at 0x000001DE672FA160>\n\n  File \"getitune\\backend\\openvino\\models\\base.py\", line 437, in _infer_compiled_model\n    out.get_any_name(): infer_request.get_tensor(out).data.copy() for out in compiled_model.outputs\n                        │             │                                      │              └ <property object at 0x000001DE3618BE20>\n                        │             │                                      └ <CompiledModel:\n                        │             │                                        inputs[\n                        │             │                                        <ConstOutput: names[x.1] shape[1,3,640,640] type: f32>\n                        │             │                                        ]\n                        │             │                                        outputs[\n                        │             │                                        <ConstOutput: names[] shape[1,39,84...\n                        │             └ <instancemethod get_tensor at 0x000001DE36182B60>\n                        └ <InferRequest:\n                          inputs[\n                          <ConstOutput: names[x.1] shape[1,3,640,640] type: f32>\n                          ]\n                          outputs[\n                          <ConstOutput: names[] shape[1,39,840...\n\nRuntimeError: Check '!get_names().empty()' failed at src\\core\\src\\descriptor\\tensor.cpp:102:\nAttempt to get a name for a Tensor without names\n\n", "record": {"elapsed": {"repr": "0:02:44.305480", "seconds": 164.30548}, "exception": {"type": "RuntimeError", "value": "Check '!get_names().empty()' failed at src\\core\\src\\descriptor\\tensor.cpp:102:\nAttempt to get a name for a Tensor without names\n", "traceback": true}, "extra": {}, "file": {"name": "base.py", "path": "app\\execution\\base.py"}, "function": "_report_progress", "level": {"icon": "❌", "name": "ERROR", "no": 40}, "line": 122, "message": "Failed: Run Quantization", "module": "base", "name": "app.execution.base", "process": {"id": 19032, "name": "job-quantize-506b7035-47ba-4d34-a1d9-fae6bc885cbd"}, "thread": {"id": 16652, "name": "MainThread"}, "time": {"repr": "2026-08-07 11:41:12.175943+02:00", "timestamp": 1786095672.175943}}}

[leoll2]

I was able to reproduce the bug in my instance, full log: yolo26-seg-quantization.log

RuntimeError: Check '!get_names().empty()' failed at src/core/src/descriptor/tensor.cpp:102:
Attempt to get a name for a Tensor without names

I run a few more experiment to see if the problem affects other scenarios too, here is what I found

• It does NOT affect regular quantization (i.e. not accuracy-aware)
• It does also affect YOLO11-Seg
• It does NOT affect other segmentation models except the Ultralytics ones
• It does NOT affect detection models, including the Ultralytics ones